Search

Chest X-ray is one of the most widely used methods for clinical diagnosis of chest diseases. In recent years, the development of deep learning technologies has driven progress in chest disease detection, but existing methods still face numerous challenges. Current research primarily focuses on detecting specific chest diseases. However, when chest X-ray images indicate multiple diseases, the diverse and complex characteristics of different disease types make it challenging to extract effective information. Additionally, the detection accuracy of small lesions remains low, which lessens the overall lesion detection rate. To address these issues, a novel network, named YOLO-CXR, is proposed in this paper for multiple disease detection, which is able to effectively locate multiple small lesions in chest X-ray images. Firstly, the proposed network enhances the YOLOv8s backbone by replacing the ordinary convolutional layers with RefConv layers to improve its feature extraction capabilities w.r.t. various diseases. Secondly, it utilizes a novel Efficient Channel and Local Attention (ECLA) mechanism to increase its sensitivity to the spatial location information of different lesions. Thirdly, to enhance its detection of small lesions, YOLO-CXR incorporates a dedicated small-lesion detection head and the Selective Feature Fusion (SFF) technique. Due to these improvements, the proposed network significantly enhances its detection of lesions at different scales and multiple small lesions in particular. Experiments conducted on the publicly available VinDr-CXR dataset demonstrate that YOLO-CXR achieves an mAP@0.5 of 0.338, a mAP@[0.5:0.95:0.05] of 0.167, and recall of 0.365, outperforming all state-of-the-art networks considered.

Tungsten coil is one of the important components of the magnetron, and its surface quality and geometry directly affect the service life of the magnetron. The focus of this article is to apply machine vision and deep learning algorithms to study a surface defect detection and geometric measurement system for tungsten coils. Firstly, in order to solve the problem of difficulty, high cost, and low efficiency in manual detection caused by small and random surface cracks on tungsten coils, this paper proposes a surface crack detection method based on deep learning technology. By constructing the dataset and training the semantic segmentation model, the surface cracks were effectively detected and identified. Secondly, a sub-pixel edge detection method based on image processing technology and the Canny algorithm is proposed to address the issue of difficulty in measuring the geometric dimensions of tungsten coils due to their high bending spiral structure. The method accurately measures the length, outer diameter, and the maximum error is only 0.023mm. The results of the verification experiment indicate that the developed system can detect an average of 70 tungsten coil samples in one minute, with sensitivity and accuracy of 99.66% and 98.52%, respectively. This system not only has high robustness and efficiency, but also reduces 95.23% of manual workload, meeting the requirements of production lines for surface defect detection and geometric dimension measurement of tungsten coils, and solving the limitations of traditional manual detection.

Wheat serves as a crucial agricultural commodity and a primary dietary staple for numerous global populations. However, it faces persistent threats from various diseases targeting wheat leaves, ultimately impacting its production. Accurate and prompt automated disease diagnosis through advanced computer vision is crucial for safeguarding wheat quality. However, the literature relied on inadequate feature selection acquired from computationally expensive backbones followed by shallow layered networks. It is resultantly limiting their capacity to recognize and prioritize diseased areas effectively. Therefore, this paper introduces an optimal features-assisted lightweight framework that integrates EfficientNet-B3 with a spatial attention (SA) mechanism to capture healthy and unhealthy patterns effectively. The proposed framework harnesses this capability to address the vital regions affected by wheat diseases via an optimized, lightweight, and attentive network. Subsequently, we thoroughly analyzed several backbone features to identify robust hyperparameters conducive to achieving our lightweight objective. Furthermore, we employed SA blocks to fortify the network, directing attention efficiently towards diseased regions. The efficacy of the proposed network is validated through comprehensive evaluations conducted on both our proposed and LWDCD2020 benchmarks. Comparative analyses with existing methods consistently showcase superiority, firmly establishing our proposed approach as a viable network for wheat disease recognition. additionally, we also collected our own dataset, namely the Wheat Disease Five Classes Classification Dataset (WD5CC), and included diverse images.

Image compression technologies such as JPEG enable efficient use of digital images. However, these technologies also degrade the quality of the image, resulting in a poor visual appearance and reduced usability. One way to solve the quality problem and achieve high efficiency in using high-quality JPEG images is to use quality improvement technologies based on artifact removal. In recent years, various methods using machine-learning-optimized neural networks have been devised to achieve high performance whereas maintaining versatility for JPEG images with various qualities. On the other hand, methods that achieve high-quality images need enormous computing costs, which is a barrier to easy use, especially when using images with large size. Given this situation, this paper proposes a method that achieves higher performance with smaller computational resources while maintaining the image quality in broader situations. The proposed method consists of a Quality Estimation Part (QE Part) that estimates the quality of the input image and Image Processing Parts (IP Parts) that process the image based on the estimated quality representation. Sequential processing is carried out by connecting multiple IP Parts in a cascade, which enables efficient and effective processing of different features. Each IP Part consists of multiple Processing Blocks, which enable effective quality improvement while maintaining efficiency. These measures enable the proposed method to achieve state-of-the-art quantitative results and better qualitative results that outperform conventional methods for images with various qualities. The code and the pre-trained models are released at https://github.com/ezumi-keio/Sequential_Processing-main.

In the consumer space, deep fakes refer to highly realistic, AI-generated images, audio, or videos that mimic real people generated by cutting-edge technologies such as Generative Adversarial Networks (GANs). In the digital age, recognizing and detecting deepfakes is a critical problem. The most common solutions for deepfake creation are those based on GANs, which can efficiently manipulate multimedia data or create from scratch. GANs comprise two neural networks, a Generator (G) and a Discriminator (D), that concurrently work during competition. The generator generates artificial data, whereas the discriminator calculates the authenticity of generated and real data. This adversarial procedure causes the generator to generate more realistic content. Identifying deep fakes produced by GANs using deep learning (DL) includes leveraging complex neural networks to detect subtle anomalies and artefacts that GANs accidentally introduce. Convolutional Neural Network (CNN) are very effective for these tasks, as they learn to discern inconsistencies and complex features in image textures, lighting, and facial features frequently missed by human eyes. This CNN model is trained on a massive database of fake and authentic images, allowing them to detect minor defects. This study presents a Deep Feature Fusion-based Fake Face Detection Generated by Generative Adversarial Networks (DF4D-GGAN) technique for Consumer Space Environment. The goal of the DF4D-GGAN technique is to detect the presence of real or deepfake images generated by DL. In the DF4D-GGAN technique, the Gaussian filtering (GF) approach is used for preprocessing the input images. Besides, the feature fusion process uses EfficientNet-b4 and ShuffleNet. Moreover, the hyperparameter selection of the DL models is performed by an improved slime mould algorithm (ISMA). Finally, an extreme learning machine (ELM) classifier has been employed to proficiently recognize real and fake images. To validate the results of the DF4D-GGAN technique, a series of simulations were made on benchmark datasets. The results stated that the DF4D-GGAN technique gains improved results over other models.

Collaborative and mobile robotics for industrial environments promise to enable autonomous and flexible production processes. However, this vision also poses significant challenges to the robotic systems, requiring them to adapt to dynamic environments and ensure human safety by leveraging continuous image streams and sophisticated data processing. Edge computing allows offloading the computational load to edge servers, communicating the image data, generated on mobile robots, over fast and reliable private 5G networks. However, there are multidisciplinary and interdependent factors that influence the reaction time of the distributed system which are not well explored in the literature for real robotic use cases, but have a significant impact on safe robotic behavior and the effectiveness of edge computing. In this work, we implement a distributed control system that offloads the image processing to measure and analyze the effect of various factors on the reaction time of the system for collaborative robotics applications. Different values for the sensing rate, image resolution and compression and quality-of-service settings are evaluated for communication and computation times as well as for the task performance. To account for the safety requirements in collaborative robotics, we add a low-level control timeout in cases of large jitter and stop the robot in cases of frequently missed detections. A push and a teleoperation experiment evaluate the reaction times in real distributed control scenarios using 5G edge computing. All experiments are implemented using ROS 2 Humble. The code and videos of the experiments are available at https://github.com/DominikUrbaniak/ros2_distributed_control_system.

Crop classification and its area estimation hold significant importance in agricultural production and management. It serves as a crucial tool for efficiently identifying various crops and contributes to the formulation of tailored strategies for crop-specific production and estimation purposes. Numerous deep learning techniques, with a particular emphasis on Convolutional Neural Networks (CNNs), have found application in the analysis of satellite data for crop classification. Nevertheless, this approach has its limitations, notably the challenge of sequential information loss, which hinders its ability to accurately extract features that rely on temporal changes over time. On the other hand, another deep learning technique, the Gated Recurrent Unit (GRU), excels at handling time series data but falls short in its ability to extract intricate patterns within the data, a strength of CNNs. This study explores the capabilities of these methodologies for crop classification in the Charsadda district of Pakistan, a region renowned for its diverse agricultural landscape dedicated to the cultivation of economically valuable crops. We introduce a dual-channel CNN-GRU feature fusion architecture, a method that merges individually extracted features from both CNNs and GRUs. The proposed framework is specifically designed for data obtained from Sentinel-2 and PlanetScope imagery, complemented by NDVI timeseries data collected from the study site. The model’s performance is assessed in comparison to the individual performance of CNN and GRU models. The models were subjected to evaluation using a variety of metrics, encompassing precision, recall, F1 Score, and overall accuracy. The outcomes reveal that the proposed model surpasses the performance of the individual models. The feature fusion-based model emerged as the top performer, achieving an overall accuracy of 96.47 percent, an F1 Score of 93.16 percent, a precision of 93.93 percent, and a recall of 92.69 percent. This research offers substantial evidence supporting the potential of the CNN-GRU feature fusion model for crop identification. By harnessing the strengths of both architectural approaches, it demonstrates its effectiveness as a valuable tool for crop classification.

Accurate estimation of path loss poses a challenge given its strong reliance on complex propagation environments that may change frequently for example due to new building constructions. Furthermore, the propagation channel at millimeter wave (mmWave) frequencies is highly susceptible to blockage, leading to notable variations in path loss values among different links at comparable distances, which renders path loss estimation based on empirical model less representative. In this paper, we tackle the challenge of accurate path loss estimation for sub-6 GHz and mmWave frequencies when working with a limited dataset of satellite images. In contrast to current literature which relies on satellite images in conjunction with supplementary models, features, or algorithms for path loss estimation, our approach involves directly estimating the path loss from satellite images. Specifically, we employ convolutional neural networks (CNNs), based on VGG-16 and ResNet-50 architectures, to fine-tune existing pretrained models, aiming to achieve accurate path loss estimation. By incorporating the appropriate layers to be fine-tuned, we are able to accelerate the training process. Furthermore, the proposed CNN path loss models yields lower prediction errors compared to the 3GPP 38.901 urban macro (UMa) empirical model by up to 5.5 dB in root mean square error (RMSE) and 4.6 dB in mean absolute error (MAE). In many instances involving different fine-tuning variants and frequencies, VGG-16 achieves lower prediction errors compared to ResNet-50. If processing times and resource consumption are a priority, ResNet-50 can be a viable choice, especially if tolerating increases in RMSE and MAE by up to 0.9 dB and 1 dB, respectively, is acceptable.

Grape cultivation faces various challenges, such as pests, management, fertilizer quality, and diseases caused by bacteria, fungi, and viruses. Notably, powdery mildew and blotches are significant diseases with different features, necessitating an accurate detection system to minimize crop losses. While traditional methods involve capturing images of diseased leaves, this research proposes a smart approach using deep learning and machine learning algorithms to analyze images taken by strategically placed cameras on farms. The research aims to design a system capable of detecting diseases that can provide information relevant to decisions, alert farmers, and allow authorized actions. Employing artificial intelligence algorithms such as support vector machines (SVM), convolutional neural networks (CNN), decision trees (DT), Naive Bayes (NB), and random forest (RF), the proposed model classifies datasets of powdery mildew, blotches, and healthy leaves when using augmented and histogram-oriented gradient (HOG) preprocessing. Following the classification of affected and healthy leaves, a stacking algorithm is used to select the optimal algorithm that provides the highest level of accuracy. The experimental results and analysis reveal that the CNN classifier outperforms others, achieving an accuracy of 96.1%. When transfer learning and fine tuning are applied to the CNN-based model, the accuracy of the model increases by 1.4% and 3.1%, respectively. SVM classification also provides a suitable level of accuracy of 96.0% for HOG augmented data.

Mean curvature is fundamentally important for various research fields, including differential geometry in mathematics, cell membrane modeling in biophysics and image processing in computer vision. However, its optimization is a challenging task. Conventionally, its optimization can be carried out by Partial Differential Equations (mean curvature flows) or solved by Augmented Lagrangian Methods. These approaches are numerically slow because they have to satisfy several numerical stability constraints. To tackle this issue, in this paper, we propose to use deep neural networks for the mean curvature optimization. Our method has several advantages. First, the proposed network has only 0.046M parameters, leading to a compact model. Second, the proposed network can be trained on low resolution images but applied on any resolution images. Third, it is flexible in the loss function design and thus can be used as a generic optimization method. To numerically valid the effectiveness and efficiency of the proposed method, several experiments are conducted. The results confirm that the proposed method achieves lower energy than traditional methods and thus establishes a new the-state-of-the-art for mean curvature optimization. The proposed method can be applied in a large range applications, where mean curvature optimization is involved.

Urine sediment analysis plays a vital role in the evaluation of kidney health. Traditional machine learning techniques approach the task of urine sediment particle detection as an image classification problem, wherein the particles are segmented based on features like edges or thresholds. However, the segmentation process for sediment particles in urine images is complex due to the inherent limitations of low contrast and weak edge characteristics. To mitigate the background noise on detection, that is, to focus more on the important part (i.e., cells), there have appeared several works that employ attention-based urinary sediment detector; however, their works did not consider the best location attention modules. This paper YOLOv7-based urinary sediment detection with attention modules. YOLOv7 is one of state-of-the-art models, and we additionally implement attention modules in the backbone of YOLOv7 so that they empower its network to enhance the feature extraction with mitigating the background noises. In experiments, we perform the proposed models on urinary sediment dataset and the results demonstrate that our proposed models outperform original YOLOv7 and a state-of-the-art urinary sediment detector in terms of recall score by 12.4% and 4.3% as well as mAP score by 7.4% and 1.6%. Our source is provided in the following link: https://github.com/Info-Sys-OU/Multi-class-Urinary-Sediment-Particles-Detection-based-on-YOLOv7-with-Attention-Modules.

Dataset quality is an integral aspect of image object detection. Consequently, poor-quality datasets are often blamed for limiting model performance. Conventional convolutional neural network models attempt to address the model performance issue and have achieved high accuracy; however, these models fail to detect objects with photometric and geometric variations in similar images. To address this object detection issue, we adopt selective image augmentation by identifying potential weak points of photometric factors, such as brightness and sharpness. A stall-housed pig dataset and a food product dataset with complex and clear backgrounds, respectively, were selected to test our hypothesis. Images from both test datasets were investigated with photometric variations ranging between −60% $\sim ~60$ % with a step size of 20% to identify the weak points of the dataset. The test results demonstrate that a photometrically balanced dataset with selective augmentation significantly outperforms the original dataset in terms of high accuracy and robustness.

The purpose of image style conversion is to transfer the style of one image to another, so that the target image retains the original content and the style of the reference image. An image style conversion technique based on generative adversarial network is proposed. This study innovatively adopts dual synthesizer and dual discriminator structure to improve the quality and efficiency of style conversion, and introduces extended convolution to enhance feature extraction. Combined with a well-designed loss function to optimize the style conversion process, a convolutional module reconstruction generator network including linear computation is added. The experimental results showed that in the training time test, the research method maintained a training time of less than 97 seconds when the number of input style types increased to 18. When conducting image style loss testing, the research method found that the image style loss value was lower compared to other techniques when the input size was 1080p and the pixel count was 10M. In the analysis of pixel loss during image style conversion, the research method shows that the pixel loss after processing virtual images is only 3.5k out of 10M, which hardly affects the expression of image content. The designed image style conversion model can accomplish the task of image style conversion with high quality and high efficiency.

To improve the prediction accuracy of monocular depth estimation networks and address issues such as edge blurring and excessive artifacts in the generated depth maps, this paper proposes a deep network architecture based on a global-local feature fusion module and a residual pooling module. The encoder utilizes a Hierarchical Transformer, while the decoder incorporates a U-Net structure model that combines multi-dimensional attention features aggregation and residual pooling. The residual pooling module facilitates better extraction of background contextual information from the feature maps to obtain more precise scene depth information. The global-local feature fusion module enables the network to learn features that encompass both global and local information. Experimental evaluations conducted on the NYU Depth V2 and KITTI datasets demonstrate that the proposed method achieves a $\delta 1$ of 0.916 on the NYU Depth V2 dataset, along with enhanced generalization ability and robustness. Furthermore, the effectiveness of each module is validated through ablation studies on the NYU Depth V2 dataset.

The efficient arrangement of patterns on a substrate surface is of paramount importance across diverse domains, including manufacturing, design, nanotechnology, and printed electronics. This paper introduces three distinct strategies for pattern arrangement within the realm of two-dimensional cutting and packing challenges. Firstly, for irregular patterns, we have developed a modified genetic algorithm that effectively determines the spatial relationships between patterns, facilitating subsequent packing processes. Secondly, we present the “MaxiRectPack” method designed for rectangular patterns. This method utilizes partitioning techniques and a variety of orientation schemes to maximize pattern packing efficiency. Lastly, we propose a technique based on normalized 2-D cross-correlation for patterns of diverse shapes, allowing precise placement while accounting for variations in substrate and pattern energy. These methodologies are adaptable, catering to irregular, rectangular, and diverse pattern shapes, offering versatile solutions for pattern placement in the context of two-dimensional cutting and packing. With a strong emphasis on resource conservation and spatial efficiency, this research equips practitioners and researchers with practical tools for optimizing material utilization and pattern arrangement on substrates.

In the domain of road inspection and structural health monitoring, precise crack identification and segmentation are essential for structural safety and disaster prediction. Traditional image processing technologies encounter difficulties in detecting cracks due to their morphological diversity and complex background noise. This results in low detection accuracy and poor generalization. To overcome these challenges, this paper introduces MixSegNet, a novel deep learning model that enhances crack recognition and segmentation by integrating multi-scale features and deep feature learning. MixSegNet integrates convolutional neural networks (CNNs) and transformer architectures to enhance the detection of small cracks through the extraction and fusion of fine-grained features. Comparative evaluations against mainstream models, including LRASPP, U-Net, Deeplabv3, Swin-UNet, AttuNet, and FCN, demonstrate that MixSegNet achieves superior performance on open-source datasets. Specifically, the model achieved a precision of 95.2%, a recall of 88.2%, an F1 score of 91.5%, and a mean intersection over union (mIoU) of 84.8%, thereby demonstrating its effectiveness and reliability for crack segmentation tasks.

Face Recognition (FR) is the technology used to identify and verify individuals based on their facial features. In recent decades, FR plays a crucial role in various sectors, including security, healthcare, banking, and criminal identification. Numerous techniques for effective FR are currently under development, ranging from appearance to hybrid approaches. Most of the existing methods offer diverse solutions to describe a face image either by focusing on specific facial features or by considering the entire face. This study explores a various range of such techniques and challenges related to FR. The existing solutions were analysed with respect to various perspectives of inputs, viz., illumination, pose variation, facial expressions, occlusions, and aging, which led to the prominent implementation of FR systems. The primary contribution of this survey lies in the comprehensive review of state-of-the-art FR techniques and deriving the taxonomy of categorizing these methods into various classes which range from appearance to hybrid approaches. Moreover, the proposed detailed study highlights the significant features used by the most recent research developed in FR, also, provide a detailed classification of image and video-based FR methods, highlighting major advancements and core processing steps for handling huge volume of datasets. Moreover, the proposed study outlines the current trends in available datasets and emphasizing their enhancements. This survey also aims to provide a valuable resource for researchers and practitioners by offering insights into the latest developments and identifying open problems that require further investigation.

Peripheral Sensory Neuropathy (PSN) affects a large proportion of individuals suffering from type 2 diabetes. To avoid ulceration and other damage to the patient’s feet, regular PSN testing, and assessment must be undertaken. Currently, the Semmes-Weinstein Monofilament Examination (SWME) is one of the most widely accepted techniques for PSN assessment. This process is time-consuming, requires special training, and is prone to errors. The number of type 2 diabetes sufferers globally is growing at alarming rates with healthcare workers under enormous pressure to continue to provide one-to-one regular care. In order to reduce the burden on existing services whilst providing the necessary care to patients, automated approaches for PSN detection provide many advantages. Importantly, with respect to an automated SWME method, there will be areas on the plantar surface where the SWM probe should not be applied i.e., areas with lesions or suspect regions. The research presented in this manuscript conducted a comprehensive analysis of different feature sets and classifiers for the task of lesion classification. Three distinct feature sets Local Binary Pattern (LBP), Mel Frequency Cepstral Coefficients (MFCC), and Scale-Invariant Feature Transform (SIFT) were evaluated across various classifiers, including Support Vector Machine (SVM), Multi-layer Perceptron (MLP), Random Forest (RF), Naïve Bayes (NB), and XGBoost. The results revealed nuanced performances across the combinations of feature sets and classifiers. While each feature set demonstrated strengths, the NB classifier applied to the LBP feature set emerged as the most notable performer with an accuracy score of 100%. This combination achieved perfect accuracy, precision, recall, and F1-score metrics, showcasing its robustness in accurately classifying lesion instances. The 5-fold cross-validation results underscored the stability of NB on the LBP feature set, with a negligible standard deviation, affirming its consistent performance across different data subsets. Additionally, the computational time complexity of 0.91 seconds highlighted its efficiency, making NB on the LBP feature set a practical and reliable choice for real-world applications. Statistical analysis using the one-way ANOVA test revealed significant differences in classifier performance across feature sets, with MFCC resulting in significantly lower accuracy compared to LBP and SIFT, which showed similar performance. The Tukey HSD post-hoc test confirmed these findings, highlighting the crucial role of feature set selection in classifier effectiveness.

The rapid development of the Internet of Things (IoT) has put great challenges on the next generation's communication technologies. Optical Camera Communication (OCC), with its advantages of the green principle, less electromagnetic interference, and harmonious integrations, can serve as a complementary, and sometimes viable alternative technology to mainstream Radio Frequency (RF)-based one. Designing efficient OCC platforms crucially involves accurate modeling of the OCC system and evaluating the efficiency of OCC-based techniques, which then stimulates related practical applications. This paper aims to provide a systematic perspective to understand OCC and to present design guidelines for OCC systems from theory to practice. First, the main parameters and analytical models for OCC basic components are presented to provide guidance on hardware parameter selection and system simulations. Next, typical implementation procedures and common performance evaluation criteria for an overall OCC system are provided. Then, detailed goal-oriented OCC enhancement solutions are summarized aiming to provide systematic guidelines for system performance enhancement. Finally, OCC designed examples are presented based on the results of our OCC experimental platform and some future research directions are discussed.

Due to the influence of rainy and foggy weather, obtaining clear images becomes more challenging, often resulting in low visibility, poor contrast, and missing detail information. To address these issues, a robust image defogging algorithm is proposed. Firstly, the input image undergoes conversion into a detailed image, with attenuation and redefinition of its three color channels. Color compensation and balance are then applied based on the principle of minimizing color loss. Secondly, the problem of image darkening is tackled through an improved atmospheric scattering model (EASM) and the dark channel a priori algorithm. The defogging results exhibit noticeable enhancements in terms of bright colors and clear details. In the natural images showcased in the paper, the proposed algorithm achieves improvements in information entropy, the fog density evaluator (FADE), and the natural image quality evaluator (NIQE) by 0.46%, 9.7%, and 12.0%, respectively, compared to the suboptimal algorithm. In the synthetic image datasets I-HAZE and O-HAZE, there are enhancements in information entropy by 0.19% and 0.76%, respectively, and in NIQE by 1.05% each, albeit slightly lower than the sub-optimal results. The structural similarity (SSIM) also sees improvements of 6.3% and 10.9% compared to the suboptimal results in FADE. These findings demonstrate the superior performance of the proposed algorithm over the latest defogging algorithms in terms of information entropy, FADE, NIQE, and SSIM, underscoring its high robustness and promising application prospects.

The analysis of aesthetic assessment is a complex and subjective task that has attracted researchers for a long time. The subjective nature of aesthetic preferences presents a significant challenge in defining and quantifying what makes images visually appealing. The current paper addresses this gap by introducing a novel methodology for quantifying and predicting aesthetic preferences in the case of interior design images. Our study combines fuzzy logic with image processing techniques. Firstly, a dataset of interior design images was collected from social media platforms, focusing on essential visual attributes such as color harmony, lightness, and complexity. Then, these features were integrated using a weighted average to compute a general aesthetic score. Our methodology considers personal color tastes when determining the overall aesthetic appeal. Initially, user feedback was collected on primary colors such as red, brown, and others to gauge their preferences. Subsequently, the image’s five most prevalent colors were analyzed to determine the preferred color scheme based on pixel count. The color scheme preference and the aesthetic score are then passed as inputs to the fuzzy inference system to calculate an overall preference score. This score represents a comprehensive measure of the user’s preference for a particular interior design, considering their color choices and general aesthetic appeal. The Two-Alternative Forced Choice (2AFC) method validated the methodology, resulting in a notable hit rate of 0.68. This study can help in fields such as art, design, advertising, or multimedia content creation, where aesthetic analysis and preference prediction are crucial. In the case of interior design, this study can help designers and professionals better understand and meet people’s preferences, especially in a world that relies heavily on digital media.

The growing use of computer vision in many fields has seen an increasing need for visual data, both real and synthetic, to support the training and testing of computer vision models. To this end, the need for high quality datasets has also increased. That said, the quality of image data for the purpose of computer vision training remains a comparatively unexplored field. There is little information on how image parameters may affect data quality, whether such parameters affect real and synthetic data equally, and how such parameters may interact with each other. The goal of this study was to test the effect of basic image data parameters on real and synthetic data and the resultant effect such parameters have on training a human action recognition model. A human action dataset with two distinct actions was selected and individual image parameters were modified to produce a total of 247 data subsets. Testing a total of 247 human action recognition models, trained on the data subsets, it was found that image parameters affect real and synthetic data differently. The impact of parameters, such as resolution and colour space, varies not just on whether the image data is real or synthesised, but also on the human action being performed. Hybrid datasets created by mixing real and synthetic data, up to 70% synthetic data, are able to produce models with similar performance levels to models trained on real data, without the negative effects of some parameters that could be observed in synthetic only datasets. Lastly, the mislabelling of training data has a large impact on action prediction accuracy, even at low percentages of mislabelled data.

Locating veins accurately is a common challenge in clinical settings, particularly with patients where veins are difficult to discern. The development of effective vein visualization technology is crucial for improving patient outcomes and minimizing procedural complications. This study aimed to design, develop, and evaluate a low-cost, high-resolution portable vein finder device. It was developed using affordable components and integrated imaging technology. The device captured images of veins in challenging patient scenarios. Four conventional image enhancement techniques were initially applied to these images. Subsequently, a novel image enhancement technique is proposed. The performance of each technique was qualitatively accessed using visual representation and histogram distribution and quantitatively assessed using the peak signal-to-noise ratio and structural similarity index measure. The proposed technique significantly outperformed the conventional methods. This indicates a superior enhancement of vein visibility, confirming the effectiveness of the proposed method. Thus, the newly developed vein finder, enhanced with the proposed image enhancement technique demonstrates significant potential for clinical use. It offers a reliable solution for vein detection in patients, potentially improving treatment accuracy.

In this paper, a novel approach to generating keys for securing color information by combining discrete wavelet transform (DWT) in conjunction with chaotic functions is presented. This procedure is based on chaotic sequences derived using initial conditions and control parameters of the Logistic map function, which are utilized to decompose in different frequency subbands by applying the DWT. These subbands construct the chaotic encoded directional phase masks (CEDPM) keys. To investigate the potential applications of the CEDPM keys, a nonlinear optical cryptosystem for securing color images has been developed. The color image to be encoded in this study is first separated into three color planes such as red (R), green (G), and blue (B) which are transformed into the phase function using a modified phase retrieval algorithm. Subsequently, for R, G, and B color planes in the encoding procedure, the obtained CEDPM keys are employed, correspondingly. This procedure helps to facilitate the creation of a real-value ciphertext for the color channels R, G, and B. These encrypted images, which correspond to the color channels R, G, and B, are positioned at the horizontal, vertical, and diagonal subbands to combine with low-frequency subbands. After that, inverse wavelet operations are carried out to produce a single covered image. This technique strengthens the system’s security by adding an additional layer. The efficacy of the color encryption approaches is confirmed using computer simulations. The results and analysis show that the proposed scheme performs satisfactorily in protecting color image components compared to the other techniques in this area.

Groundnut is a crucial crop on a global scale and is facing considerable decrease in yield due to leaf diseases. This makes it essential for the prompt identification of these diseases to ensure continuous agricultural productivity. In this research article, LeafNet, a novel architecture, is introduced for detecting six major classifications of groundnut leaf, including diseases such as Early Rust, Early Leaf Spot, Nutrition Deficiency, Rust, Late Leaf Spot, as well as Healthy Leaf, from a dataset comprising 10,361 images. The performance of the proposed architecture is evaluated through a variety of subjective and objective assessment techniques to validate its efficiency. The robust performance of LeafNet can be attributed to utilizing residual networks and weight initialization techniques within its ensembling process. The proposed architecture undergoes comparison with various neural network architectures introduced in state-of-the-art techniques, demonstrating an exceptional accuracy. The proposed architecture achieved a test accuracy of 97.225%, precision of 97.365%, recall of 97.225%, F1-score of 97.225%, and MCC of 96.700%. To assess the adaptability and generalizability of the proposed architecture, it is evaluated with various leaf disease datasets along with groundnut leaf diseases. This research makes a substantial contribution to agricultural science by presenting a robust deep learning-based architecture for precise identification of groundnut leaf diseases. The potential impact of this method on agricultural practices is significant, facilitating timely disease management and reinforcing agricultural sustainability.

In the context of food science and engineering, the in vitro chewing effect on food bolus formation is a critical area of research that explores the mechanical and textural properties of ingested materials. This article presents a pioneering approach to assess the in vitro chewing impact on food bolus formation using the gray level co-occurrence matrix (GLCM) image analysis technique. As technological advancements lead to the development of mastication robots, the need for evaluating in vitro chewed food bolus has grown. To address this challenge, a case study is conducted. The study's objectives encompass utilizing GLCM to determine the in vitro chewing cycle phase, analyzing texture features, and investigating chewing trajectory differences for beef and plant-based burger patties. Applying GLCM as a methodology, the research quantitatively analyzes textural features of food bolus formations under controlled in vitro chewing conditions. The outcomes reveal distinct differences between beef and plant-based samples through GLCM parameters. Significantly, the study identifies a consistent trend across various scenarios, indicating an increase in energy and homogeneity and a decrease in dissimilarity with an increasing number of in vitro chewing cycles. This investigation offers valuable insights into the dynamic relationship between chewing cycles and textural features in the oral processing of beef and plant-based burger patties.

Edge-preserving smoothing is crucial in image processing for removing noise and fine textures while maintaining significant structures. This paper focuses on filter-based methods due to their computational efficiency and ease of implementation. Edges contain essential information defining object boundaries and texture details, characterized by both intensity and scale. Traditional filters, such as the bilateral, domain transform, and guided filters, primarily rely on edge intensity without the ability to adjust scale. This limitation prevents them from effectively smoothing small-scale textures while preserving significant structures. To address this limitation, we propose an edge-preserving smoothing filter that enables real-time control of both edge intensity and scale. Our method introduces a novel metric based on the variance of pixel values within patches to quantitatively assess regional flatness at a specific scale. The fundamental idea is to smooth patches at a specific scale to remove smaller-scale details while preserving larger-scale structures. Each pixel is assigned a weighted average of the smoothed results from multiple overlapping patches, with the weights determined by the inverse of the patch variances. This approach allows adaptive filtering that effectively smooths textures while preserving significant edges. Experimental comparisons with conventional methods demonstrate that our proposed filter efficiently removes textures and noise while preserving significant edges. By providing immediate visual feedback, our method allows rapid adjustments of both scale and intensity, making it suitable for real-time applications. Future work will focus on adaptive scale control to develop a texture suppression filter adaptable to diverse image structures and textures.

Point source detection algorithms play a pivotal role across diverse applications, influencing fields such as astronomy, biomedical imaging, environmental monitoring, and beyond. This article reviews the algorithms used for space imaging applications from ground and space telescopes. The main difficulties in detection arise from the incomplete knowledge of the impulse function of the imaging system, which depends on the aperture, atmospheric turbulence (for ground-based telescopes), and other factors, some of which are time-dependent. Incomplete knowledge of the impulse function decreases the effectiveness of the algorithms. In recent years, deep learning techniques have been employed to mitigate this problem and have the potential to outperform more traditional approaches. The success of deep learning techniques in object detection has been observed in many fields, and recent developments can further improve the accuracy. However, deep learning methods are still in the early stages of adoption and are used less frequently than traditional approaches. In this review, we discuss the main challenges of point source detection, as well as the latest developments, covering both traditional and current deep learning methods. In addition, we present a comparison between the two approaches to better demonstrate the advantages of each methodology.

This work introduces Orthogonal Lattice Universal Wavelet Unit, a novel trainable wavelet unit to enhance image classification and anomaly detection in convolutional neural networks by reducing information loss during pooling. The unit employs an orthogonal lattice structure, relaxing the zero-at- $\pi $ condition and decreasing the number of trainable wavelet coefficients. This innovation is a key novelty of the work. The unit modifies convolution, pooling, and down-sampling operations. Implemented in residual neural networks with 18 layers, it improved detection accuracy on CIFAR10 (by 2.67%), ImageNet1K (by 1.85%), and the Describable Textures dataset (by 9.52%), showcasing its advantages in detecting detailed features. Similar gains were seen in the implementations for residual neural networks with 34 layers and 50 layers. For anomaly detection in hazelnut images on the MVTec Anomaly Detection dataset, the proposed method achieved a segmentation area under the receiver operating characteristic curve of 97.15% and better anomaly localization. The method excels in detecting detailed features, despite increased trainable parameters from using one-layer fully convolutional networks for feature combination.

Medical ultrasound imaging involves the use of high-frequency sound waves to produce images of various body parts. A transducer generates these sound waves, which traverse through bodily tissues, providing measurements of soft tissue and organ dimensions, shapes, and consistencies. The quality of an ultrasound image depends on the frequency of the transducer used. Higher-frequency transducers yield better resolution, but they are limited in their ability to penetrate deeply into the body due to their shorter wavelengths, making them more susceptible to absorption. Lower-frequency transducers can penetrate deeper because they have longer wavelengths, although their resolution isn’t as sharp as that of high-frequency transducers. The primary drawback associated with ultrasound imaging is the introduction of noise during the signal processing stage, which can lead to images that are challenging to interpret. Efficient medical image processing plays a pivotal role in improving the quality and utility of ultrasound imaging for medical applications, enhancing image comprehension, and aiding in accurate diagnoses. One critical aspect of the pre-processing stage in medical image analysis, particularly in ultrasound images, is speckle noise reduction. To improve the analysis and diagnosis in various applications, it has become essential to employ software tools for speckle noise removal. In this paper we propose hybrid filter, which is combination of filters to be used to remove noise along with Convolutional Neural Networks. Experiments were performed at different speckle variances and the results showed that hybrid filter performed well when the speckle variance is between 0.1 and 0.3, while DnCNNL5 performed well when the speckle variance was high (between 0.5 and 0.9).

Coding algorithms are usually designed to faithfully reconstruct images, which limits the expected gains in compression. A new approach based on generative models allows for new compression algorithms that can reach drastically lower compression rates. Instead of pixel fidelity, these algorithms aim at faithfully generating images that have the same high-level interpretation as their inputs. In that context, the challenge becomes to set a good representation for the semantics of an image. While text or segmentation maps have been investigated and have shown their limitations, in this paper, we ask the following question: do powerful foundation models such as CLIP provide a semantic description suited for compression? By suited for compression, we mean that this description is robust to traditional compression tools and, in particular, quantization. We show that CLIP fulfills semantic robustness properties. This makes it an interesting support for generative compression. To make that intuition concrete, we propose a proof-of-concept for a generative codec based on CLIP. Results demonstrate that our CLIP-based coder beats state-of-the-art compression pipelines at extremely low bitrates (0.0012 BPP), both in terms of image quality (65.3 for MUSIQ) and semantic preservation (0.86 for the Clip score).

Data augmentation improves machine learning model performance by diversifying training data. Initially, manual selection of augmentation techniques was required; however, AutoAugment automated this process. However, recent studies identified limits in the search space of AutoAugment, which hindered performance gains. In this study, we introduced a novel method called Blend AutoAugment that utilizes linear blending with random weights to enhance the performance of image classification in deep learning. Our method consists of two stages: 1) policy search and 2) image blending. In the policy search stage, optimal policies are determined by employing reinforcement learning to explore a search space, comprising the types of operations, their application probabilities, and their magnitudes. In the image blending stage, multiple optimal policies discovered in the previous stage are selected to generate augmented images, and the generated images are blended using random weights. Performance comparison experiments were carried out using the Wide-ResNet40-2 and Wide-ResNet28-10 models on the CIFAR-10, CIFAR-100, CIFAR-10-C, and CIFAR-100-C datasets. In addition, the proposed method was compared with existing methods. Notably, the proposed method exhibited higher accuracy and robustness than those of existing methods. Furthermore, it was applicable to various existing automatic data augmentation methods. In summary, this study introduced an advanced automatic data augmentation method that expands the previously limited search space and utilizes image blending to achieve better accuracy and robustness than conventional methods. Our source code is available at https://github.com/Sarrzae/BAA.

Hemoglobin is the iron containing protein in red blood cells which carries oxygen from lungs to rest of the body tissues. Accurate measurement of hemoglobin is essential for diagnosing anemia, a condition characterized by a deficiency of red blood cells. This measurement is particularly vital before initiating blood transfusions for thalassemia patients. Non-invasive estimation of hemoglobin levels can be achieved through photoplethysmography (PPG)-based methods. PPG is an optical method to measure blood volume changes in successive heart beats. PPG signals can be obtained from fingertip videos using a light source and a photodetector. SmartphonePPG utilizes a smartphone’s flashlight as a light source and its camera as a photodetector to acquire PPG signals, offering an affordable and portable point-of-care tool. Despite the ubiquity of smartphones, signals from their cameras often contain noise, making feature selection from PPG characteristics challenging. While PPG-based methods are invaluable, the lack of real-world datasets poses a significant challenge in maximizing the benefits of PPG technology. In this paper, we introduce a dataset comprising 1-minute fingertip video recordings from 150 anemic patients, obtained using a smartphone’s camera. The dataset, publicly accessible for research purposes (https://forms.gle/LB4qn81ZMqEuy3V27), covers an age range of 6 months to 32 years, with diverse hemoglobin values (4.3 gm/dL - 12.4 gm/dL). Utilizing this dataset, we propose a deep learning-based technique employing the ResNet-18 architecture to estimate hemoglobin levels. This approach eliminates the need for manual feature extraction and selection from PPG signals, overcoming a limitation in existing smartphone PPG-based hemoglobin estimation systems. Our model achieves a hemoglobin level estimation with an RMSE of 0.81-1.39 when compared with the gold standard laboratory method, Complete Blood Count (CBC) test reports. In contrast, HemaApp, a state-of-the-art research utilizing a machine learning-based classifier (SVM), yields an RMSE of 1.7 on our dataset. The accuracy and simplicity of our model position it as a promising alternative to existing non-invasive hemoglobin level estimation methods.

Breast cancer is the most commonly diagnosed cancer among women, globally. The occurrence and fatality rates are high for breast cancer compared to other types of cancer. The World Cancer report 2020 points out early detection and rapid treatment as the most efficient intervention to control this malignancy. Histopathological image analysis has great significance in early diagnosis of the disease. Our work has significant biological and medical potential for automatically processing different histopathology images to identify breast cancer and its corresponding grade. Unlike the existing models, we grade breast cancer by including both local and global features. The proposed model is a hybrid multi-class classification model using depth-wise separable convolutional networks and transformers, where both local and global features are considered. In order to resolve the self-attention module complexity in transformers patch merging is performed. The proposed model can classify pathological images of public breast cancer data sets into different categories. The model was evaluated on three publicly available datasets, like BACH, BreakHis and IDC. The accuracy of the proposed model is 97.800 % on the BACH dataset, 98.130 % on BreakHis dataset and 98.320 % for the IDC dataset.

This study aims addresses the challenge of gesture recognition in smart sports stadiums through the fusion enhancement of YOLOv5 and Copula Bayesian Classifier algorithm. By leveraging FasterNet for YOLOv5 transformation, integrating the Convolutional Block Attention Module (CBAM) as an attention mechanism, and merging it with the Copula Bayesian Classifier, we propose a tandem model for robust gesture recognition in the complex environments of sports venues. Extensive training on a substantial dataset yields an efficient and accurate gesture recognition model. Our findings showcase the effectiveness of the proposed algorithm, achieving a remarkable accuracy rate of 99.2% in identifying gesture categories swiftly. This advancement holds significant implications for future human-computer interaction in smart sports stadiums, offering a more seamless and intelligent mode of interaction for both spectators and athletes. Ultimately, it enhances user experience and management efficiency within smart sports venues.

In its long history, chum salmon has been propagated by hatching and stocking. Age assessment is necessary to monitor the migration status of chum salmon, and it is desirable to increase the number of assessments in order to obtain precise and accurate knowledge about salmon as a resource. This study introduced an automated age assessment system that uses information technologies, such as artificial neural networks and image processing, to improve efficiency compared to manual assessment. Specifically, we first developed a method for extracting scale regions from scale replica sample images and creating a database. We then used a resting zone detection technique based on semantic segmentation, allowing us to automate evaluation using scaled images as part of the age assessment method. Age assessment from resting zone images was realized by an image processing method based on the skills of the staff of the Japanese Fisheries Research Institute, such as an age counting method using circular structures. Experimental results show that our proposed method outperforms other approaches; salmon age assessment was accurate to within an error of 0.5 years. Our method encompasses automatic data processing of scale images, resting zone detection, and age assessment, and will contribute to the efficiency of the Fisheries Research Institute in terms of both work efficiency and salmon data research.

Against the backdrop of the pervasive global challenge of cancer, with particular emphasis on lung cancer (LC), this study centers its investigation on the critical realm of early detection leveraging artificial intelligence (AI) within the domain of histological image analysis. Through the fusion of DenseNet201 with color histogram techniques, a novel hybrid feature set emerges, engineered to elevate classification accuracy. The comprehensive evaluation encompasses eight diverse machine learning (ML) algorithms, spanning from K-Nearest Neighbors (KNN) to Support Vector Machines (SVM), including notable contenders such as LightGBM (LGBM), CatBoost, XGBoost, decision trees (DT), random forests (RF), and multinomial naive Bayes (MultinomialNB). This rigorous examination illuminates a distinguished model, achieving a remarkable accuracy rate of 99.683% on the LC25000 dataset. The extension of this methodology to breast cancer detection, utilizing the BreakHis dataset, yields a commendable accuracy rate of 94.808%. These findings underscore the transformative potential of AI in the intricate landscape of histopathological analysis, positioning it as a pivotal force in advancing diagnostic capabilities. A meticulous comparative analysis not only underscores the merits but also elucidates the limitations of existing AI applications in medical imaging, thereby charting a roadmap for future refinements and clinical deployments. Consequently, continued research in AI within clinical settings is advocated, with the ultimate aim of fortifying early cancer diagnosis and subsequently enhancing patient outcomes through judicious therapeutic interventions.

People are preoccupied with their own duties in today’s environment, making it tough to oversee and care for disabled people. Paralyzed and disabled people, on the other hand, have a powerful desire to move around freely. Different initiatives have been taken in the past to improve and preserve the self-esteem of such people, and various technologies have also been created to assist them in a better way. The primary advantage of a vision-based autonomous wheelchair is that it can provide greater independence to users with limited mobility. The methodology of other smart wheelchairs varies depending on the specific technology used. Some smart wheelchairs may use sensors to detect obstacles and provide feedback to the user, while others may incorporate GPS and other navigation technologies to assist with indoor and outdoor navigation. Our study proposes a revolutionary implementation of an autonomous system for fully handicapped people, allowing them to drive wheelchairs using just their eyes. The wheelchair’s orientation will be determined by the orientation of the eye. The camera will follow the movement of the eyes. Furthermore, sensors will be installed on the front side of the wheelchair to identify any obstacles along its path. The calibration of the camera with the human eye, without obstructing its vision, is the most challenging task in this research, besides controlling the wheels of the wheelchair.

Luojia 3-01 is the world's first intelligent remote sensing satellite, equipped with various imaging modes including video, frame-pushing, and scan-pushing. It boasts submeter level multimode optical imaging capabilities, on-orbit intelligent processing, and real-time intra-satellite and satellite-to-ground data transmission. Throughout the satellite's design and development, we have established the real-time intelligent information service architecture for Luojia 3-01, marking a new paradigm in on-orbit processing and real-time services for intelligent remote sensing satellites. We have proposed the on-orbit real-time processing architecture of Luojia 3-01, addressing the challenges of limited computational and storage resources, particularly in on-orbit processing of vast volumes of remote sensing data, including core algorithmic bottlenecks in “correction-extraction-compression.” A novel intelligent remote sensing satellite system has been developed, featuring multimode imaging, an open platform, intelligent processing, and satellite-to-ground interconnectivity, which significantly reduces the response time of remote sensing services to less than 8 min, thereby shortening the cycle from data acquisition to intelligent information service. This innovation spearheads a technological leap in remote sensing satellite services from data to information, from post-event to real-time, and from professional to widespread public applications, laying a solid foundation for the popularization and commercialization of China's intelligent remote sensing satellites.

Masks play a crucial role in preventing respiratory diseases and have diverse applications in national public health and industrial safety. Efficient mask-wearing detection systems are essential for ensuring accuracy and real-time performance. To overcome the challenges of extensive model calculations, extensive parameter volume, and complex hardware deployment in the current mask-wearing detection system, a lightweight mask detection model with improved YOLOv5 is proposed. Firstly, this study proposes a new lightweight network—EMA-FasterNet, as the backbone network of YOLOv5, which reduces the computation while preserving the information of each channel. Secondly, using Depthwise Separable Convolutions (DepthSepConv) to replace some C3 modules in the Neck further compresses the model volume, parameters, and computation. Finally, to prevent the detection of missing objects due to the removal of overlapping candidate boxes, use Soft Non-Maximum Suppression (NMS) to replace NMS. Compared with YOLOv5s, the proposed YOLOv5-S2C2 has high mAP in both Dataset I and Dataset II, while the parameters and computation of the model are reduced by about 56% and 57.0%, respectively. The volume is about 44% of the original model and the inference speed on GPU is improved by about 30%. These improvements show that the proposed model achieves a lightweight design and excellent real-time performance. It is also well suited for efficient use on hardware with limited computational resources.

Agriculture is the ultimate imperative and primary source of origin to furnish domestic income for multifarious countries. The disease caused in plants due to various pathogens like viruses, fungi, and bacteria is liable for considerable monetary losses in the agriculture corporation across the world. The security of crops concerning quality and quantity is crucial to monitor disease in plants. Thus, recognition of plant disease is essential. The plant disease syndrome is noticeable in distinct parts of plants. Nonetheless, commonly the infection is detected in distinct leaves of plants. Computer vision, deep learning, few-shot learning, and soft computing techniques are utilized by various investigators to automatically identify the disease in plants via leaf images. These techniques also benefit farmers in achieving expeditious and appropriate actions to avoid a reduction in the quality and quantity of crops. The application of these techniques in the recognition of disease can avert the disadvantage of origin by a factious selection of disease features, extraction of features, and boost the speed of technology and efficiency of research. Also, certain molecular techniques have been established to prevent and mitigate the pathogenic threat. Hence, this review helps the investigator to automatically detect disease in plants using machine learning, deep learning and few shot learning and provide certain diagnosis techniques to prevent disease. Moreover, some of the future works in the classification of disease are also discussed.

Understanding liquid spray is essential for spray applications, including but not limited to designing fuel-efficient engines. Due to the challenges involved in collecting real-world liquid spray images, a synthetic liquid spray was generated using fluid simulation based on the atomization of the liquid jet. Semantic segmentation was chosen to analyze the liquid spray, as it reflects the precise location of the objects in the image. This paper presents a workflow to train a U-Net with a small sample (only 24 training images) dataset under the constraint that no ground truth is provided. An image is selected from the generated images of liquid spray and edited by randomly masking some objects. After the chosen image is annotated, a data augmentation technique that includes rotation and Gaussian smoothing is applied, resulting in 24 images available as the training set. The RGB original-sized images are fed to U-Net for training. Due to how the liquid spray images are obtained in the real world, Gaussian smoothing is explored as the inductive bias. Gaussian smoothing is incorporated between the convolutional layers of U-Net to enhance its feature extraction ability. The experiment results showed that the segmentation output improved when smoothing was incorporated into the U-Net. By visualizing the convolutional feature map of trained U-Net, we discover that smoothing makes convolution less biased to texture information. Going through this workflow, the trained U-Net is found to generalize well to the test images despite learning from few samples. Code is available at https://github.com/lynerlwl/spray-unet

To improve the effect of table tennis tactical analysis, this study proposed a table tennis video tactical analysis method based on image fuzzy edge recognition algorithm. This algorithm can effectively recognize edges in images to improve recognition accuracy and reduce errors and rejection rates. In addition, this research combines the methods of target tracking and trajectory estimation to build a table tennis video game tactical analysis platform. The experimental data show that the recognition accuracy of the proposed fuzzy edge recognition algorithm is 98.1%, and the error and rejection rates are less than 1.6%. Compared with the image edge detection algorithm based on interval value fuzzy, the proposed algorithm has better performance. When the positioning error is 10%, the precision of the table tennis target tracking algorithm can reach 94.3%, which is 9.3% higher than that of the direct linear transform algorithm. The table tennis video game tactical analysis platform can analyze the velocity of ping-pong ball according to the video, and records the droppoint of ping-pong ball, which helps the players to develop more effective tactics and strategies to improve their competitive level and achievements.

Acute lymphoblastic leukemia (ALL), one of the prevalent types of carcinogenic disease, has been seen a deadly illness exposing numerous patients across the world to potential threats of lives. It impacts both adults and children providing a narrow range of chances of being cured if diagnosed at a later stage. A definitive diagnosis often demands highly invasive diagnostic procedures thereby proving time-consuming and expensive. Peripheral Blood Smear (PBS) images have been playing a crucial role in the initial screening of ALL in suspected individuals. However, the nonspecific nature of ALL poses a substantial challenge in the analysis of these images thus leaving space for misdiagnosis. Aiming at contribute to the early diagnoses of this life-threatening disease, we put forward automated platform for screening the presence of ALL concerning its specific subtypes (benign, Early Pro-B, Pro-B and Pre-B) using PBS images. The proposed web based platform follows weighted ensemble learning technique using a Residual Convolutional Neural Network (ResNet-152) as the base learner to identify ALL from hematogone cases and then determine ALL subtypes. This is likely to save both diagnosis time and the efforts of clinicians and patients. Experimental results are obtained and comparative analysis among 7 well-known CNN Network architectures (AlexNet, VGGNet, Inception, ResNet-50, ResNet-18, Inception and DenseNet-121) is also performed that demonstrated that the proposed platform achieved comparatively high accuracy (99.95%), precision (99.92%), recall (99.92%), F1-Score (99.90%), sensitivity (99.92%) and specificity (99.97%). The promising results demonstrate that the proposed platform has the potential to be used as a reliable tool for early diagnosis of ALL and its sub-types. Furthermore, this provides references for pathologists and healthcare providers, aiding them in producing specific guidelines and more informed choices about patient and disease management.

With the continuous acceleration of urbanization, the parking problem is becoming increasingly serious. How to better manage parking resources has become an urgent problem to be solved in urban development. In this context, according to the historical data and real-time video data collected by the parking camera, this paper proposes an algorithm for parking space detection and state recognition. Through image preprocessing, region of interest selection, Hough line detection, and parking information recognition of the input test image, an intelligent parking space detection model is constructed, which improves the utilization rate of parking space and reduces the management cost. On this basis, according to the free parking space data obtained by the detection algorithm, a short-term demand prediction algorithm for on-road parking based on Convolutional Neural Network (CNN) and Long Short-Term Memory Neural Network (LSTM) was proposed. Through the preprocessing of input parking space data, time vector transformation, data separation, model training, and prediction, the parking demand data is predicted and analyzed. By comparing the prediction results of multiple models, it was found that the CNN-LSTM prediction model had the best model stability and goodness of fit, the lowest Mean Error (MAE) and Root Mean Square Error (RMSE), the errors of working days were 13.301 and 21.156, and the errors of rest days were 12.573 and 20.739, respectively. It shows that CNN-LSTM can effectively capture the time and spatial feature information of parking lot free parking space data, and the prediction accuracy is good, which can be used to predict the number of free parking spaces in parking lots.

Real-time Monte Carlo denoising aims to denoise a 1spp-rendered image with a limited time budget. Many latest techniques for real-time Monte Carlo denoising utilize temporal accumulation (TA) as a pre-processing to improve the temporal stability of successive frames and increase the effective spp. However, existing techniques using TA used to suffer from significant performance degradation when TA does not work well. In addition, they have the disadvantage of deteriorating performance in dynamic scenes because pixel information of the current frame cannot be sufficiently utilized due to the pixel averaging effect between temporally adjacent frames. To solve this problem, this paper proposes a framework that utilizes both 1spp images and temporally accumulated 1spp (TA-1spp) images. First, the multi-scale kernel prediction module estimates kernel maps for filtering 1spp images and TA-1spp images, respectively. Then, the filtered images are properly fused so that the two advantages of 1spp and TA-1spp images can create synergy. Also, the remaining noise is removed through the refinement module and fine details are reconstructed to improve the model flexibility, beyond using only the kernel prediction module. As a result, we achieve better quantitative and qualitative performance at 39% faster than state-of-the-art (SOTA) real-time Monte Carlo denoisers.

Pomegranates are nutrient-rich fruits renowned for their vibrant ruby-red seeds and antioxidant properties. With a rich history rooted in various cultures, pomegranates have gained widespread popularity for their distinct flavor and potential health benefits. Timely detection and understanding of the growth stages of pomegranates can facilitate optimized resource allocation, targeted interventions, and efficient crop management. Additionally, early detection contributes to maximizing crop yield, ensuring product quality, and mitigating potential risks such as diseases and pest infestations. The primary goal of the present study is the early detection of pomegranate growth stages using an efficient approach. We conducted our experiments using standard image data of the pomegranate growth stages, comprising 5857 files categorized into five classes: Bud, Early-Fruit, Flower, Mid-growth, and Ripe. We propose a transfer learning-based CRnet approach to capture spatial features from pomegranate images depicting the five stages of pomegranate growth. The extracted spatial features serve as inputs for the random forest method, resulting in the creation of a new probabilistic feature set. These new probabilistic features assist the proposed model in performing the detection of pomegranate growth stages. To evaluate performance, we implemented state-of-the-art image classification techniques, including a Convolutional Neural Network (CNN), K-Neighbors Classifier (KNC), Gaussian Naive Bayes (GNB), and Logistic Regression (LR). To ensure the accuracy of applied machine learning methods, we utilized a hyperparameter optimization approach and a k-fold-based cross-validation technique. Additionally, computational complexity is determined. Extensive analysis of research results shows that by using the proposed features, the random forest model outperformed state-of-the-art methods with a high accuracy of 98% in predicting pomegranate growth stages. Our proposed scheme has the potential for the timely detection of pomegranate growth stages, assisting farmers in maximizing crop yield and mitigating potential risks.

The most unexpected and fatal kind of cancer, melanoma, has been on the rise in its spread to different parts of the body. Early detection can significantly reduce its fatality rate. The primary stages of melanoma for its identification by the unaided eye are a difficult task that demands extensive training and understanding. Due to a lack of qualified dermatologists, a computerized and automated method is required to correctly detect melanoma. This study achieved this feat through a proposed system that can effectively detect and classify melanoma as benign or malignant. The process begins with image template matching by using normalized cross-correlation technique induction to mark the infected area of skin lesion as the region of interest (ROI) from ISIC datasets 2017, 2019, and 2020 dermoscopic images. Our novel model dynamically calculated number of clusters is assigned to the $k$ -means clustering algorithm, and ROI is extracted. Histogram equalization is applied to the output image for contrast enhancement. Hu Moment method is implemented for shape classification and part recognition from the segmented image. GLCM-based Haralick feature extractor is used in the proposed system to extract textural features generating the feature vector from the segmented skin lesion. It leads the classifier to identify skin lesions as cancerous or non-cancerous. Random Forest, Decision Tree, Support Vector Machine (SVM), Gaussian Naïve Bayes, Logistic Regression, and K-nearest neighbor (KNN) Classifiers are used for classification. The KNN, SVM, and Random Forest was successful in attaining the highest accuracy of 99.29%, 99.38%, and 99.46% on the given dermoscopic images datasets ISIC-2019, ISIC-2020 and ISIC-2017 respectively. The proposed method, the normalized cross correlation-based k-means clustering model, was found to be more robust and accurate than existing methods and incorporates much more feature information from the images.

Current carotid atherosclerosis diagnostic protocols do not feature techniques that would allow for early or frequent medical examinations, leaving a significant number of asymptomatic carotid stenosis cases undetected and often leading to strokes. The key challenge is that current diagnostics are highly operator-dependent. In this work we used idealised biological models to demonstrate a new rapid, potentially inexpensive and operator-independent diagnostic method, aimed at detecting whether a stenosis exists, rather than seeking to be accurately quantifying or localising it. An array of electrodes was used to obtain sequential bioimpedance values over the skin, through a novel scanning technique, covering an area over the artery of interest. FEM simulations, verified through in-vitro experiments on gelatine phantoms, were used to validate the method. The final results, obtained through image processing algorithms, were in the form of planar bio-impedance maps and were successful both in identifying arterial features and detecting the presence of stenoses of different sizes, and the overall accuracy of detecting stenoses in a vessel without bifurcation is 97.78% and with bifurcation is 91.11%. The results could also be used to indicate the artery’s relative orientation to the sensor, eliminating the need for manual alignment by a specialist operator. Therefore, this method shows promise for routine medical examination, either in primary care, or even at home, to indicate whether a patient would require further, more detailed examinations at a specialist clinic.

Dorsal hand vein (DHV) recognition is a burgeoning biometric technology that has recently garnered considerable attention. This article uses image processing and deep learning to present a novel DHV recognition approach. It involves detecting and identifying the unique patterns present in the DHV. The proposed system begins with the preprocessing mechanism that is applied to enhance the quality of the acquired images, including contrast enhancement and noise reduction, by using some filters such as Median and Contrast Limited Adaptive Histogram Equalization (CLAHE). Next, a deep learning model, such as a convolutional neural network (CNN), is employed to automatically abstract discriminative features from the preprocessed vein images. The empirical outcomes prove the influence and reliability of the proposed technique for vein recognition, making it a promising solution for biometric authentication systems. Compared with traditional CNN, the proposed approach shows good accuracy and classification rate results. The suggested model achieved a high recognition rate accuracy, recall, precious, and f-score of 99.7%,97%,96%, and 96%, respectively, and a recognition time of about 1283.45 s. To enrich the model’s capability for feature recognition and reduce recognition time, decrease the intricacy of learning and the connectivity CNN structure, an alternative approach based on Restricted Boltzmann Machines (RBM) was assessed. This strategy exhibits superior accuracy in comparison to other contemporary algorithms. The proposed RBM achieved a high recognition rate accuracy, recall, precious, and f-score of 99.9%,99%,99%, and 99%, respectively, and a recognition time of about 137.235s.