Search

Scene text detection is an important research field in computer vision, playing a crucial role in various application scenarios. However, existing scene text detection methods often fail to achieve satisfactory results when faced with text instances of different sizes, shapes, and complex backgrounds. To address the challenge of detecting diverse texts in natural scenes, this paper proposes a multi-scale natural scene text detection method based on attention feature extraction and cascaded feature fusion. This method combines global and local attention through an improved attention feature fusion module (DSAF) to capture text features of different scales, enhancing the network's perception of text regions and improving its feature extraction capabilities. Simultaneously, an improved cascaded feature fusion module (PFFM) is used to fully integrate the extracted feature maps, expanding the receptive field of features and enriching the expressive ability of the feature maps. Finally, to address the cascaded feature maps, a lightweight subspace attention module (SAM) is introduced to partition the concatenated feature maps into several sub-space feature maps, facilitating spatial information interaction among features of different scales. In this paper, comparative experiments are conducted on the ICDAR2015, Total-Text, and MSRA-TD500 datasets, and comparisons are made with some existing scene text detection methods. The results show that the proposed method achieves good performance in terms of accuracy, recall, and F-score, thus verifying its effectiveness and practicality. [ABSTRACT FROM AUTHOR]

This work presents a novel approach to enhancing iris recognition systems through a two-module approach focusing on low-level image preprocessing techniques and advanced feature extraction. The primary contributions of this paper include: (i) the development of a robust preprocessing module utilizing the Canny algorithm for edge detection and the circle-based Hough transform for precise iris extraction, and (ii) the implementation of Binary Statistical Image Features (BSIF) with domain-specific filters trained on iris-specific data for improved biometric identification. By combining these advanced image preprocessing techniques, the proposed method addresses key challenges in iris recognition, such as occlusions, varying pigmentation, and textural diversity. Experimental results on the Human-inspired Domain-specific Binarized Image Features (HDBIF) Dataset, consisting of 1892 iris images, confirm the significant enhancements achieved. Moreover, this paper offers a comprehensive and reproducible research framework by providing source codes and access to the testing database through the Notre Dame University dataset website, thereby facilitating further application and study. Future research will focus on exploring adaptive algorithms and integrating machine learning techniques to improve performance across diverse and unpredictable real-world scenarios. [ABSTRACT FROM AUTHOR]

Object detection is one of the core technologies for autonomous driving. Current road object detection mainly relies on visible light, which is prone to missed detections and false alarms in rainy, night-time, and foggy scenes. Multispectral object detection based on the fusion of RGB and infrared images can effectively address the challenges of complex and changing road scenes, improving the detection performance of current algorithms in complex scenarios. However, previous multispectral detection algorithms suffer from issues such as poor fusion of dual-mode information, poor detection performance for multi-scale objects, and inadequate utilization of semantic information. To address these challenges and enhance the detection performance in complex road scenes, this paper proposes a novel multispectral object detection algorithm called MRD-YOLO. In MRD-YOLO, we utilize interaction-based feature extraction to effectively fuse information and introduce the BIC-Fusion module with attention guidance to fuse different modal information. We also incorporate the SAConv module to improve the model's detection performance for multi-scale objects and utilize the AIFI structure to enhance the utilization of semantic information. Finally, we conduct experiments on two major public datasets, FLIR_Aligned and M3FD. The experimental results demonstrate that compared to other algorithms, the proposed algorithm achieves superior detection performance in complex road scenes. [ABSTRACT FROM AUTHOR]

Abnormal valve positions can lead to fluctuations in the process industry, potentially triggering serious accidents. For processes that frequently require operational switching, such as green chemical processes based on renewable energy or biotechnological fermentation processes, this issue becomes even more severe. Despite this risk, many plants still rely on manual inspections to check valve status. The widespread use of cameras in large plants now makes it feasible to monitor valve positions through computer vision technology. This paper proposes a novel real-time valve monitoring approach based on computer vision to detect abnormalities in valve positions. Utilizing an improved network architecture based on YOLO V8, the method performs valve detection and feature recognition. To address the challenge of small, relatively fixed-position valves in the images, a coord attention module is introduced, embedding position information into the feature channels and enhancing the accuracy of valve rotation feature extraction. The valve position is then calculated using a rotation algorithm with the valve's center point and bounding box coordinates, triggering an alarm for valves that exceed a pre-set threshold. The accuracy and generalization ability of the proposed approach are evaluated through experiments on three different types of valves in two industrial scenarios. The results demonstrate that the method meets the accuracy and robustness standards required for real-time valve monitoring in industrial applications. [ABSTRACT FROM AUTHOR]

Currently, the intelligent defect detection of massive grid transmission line inspection pictures using AI image recognition technology is an efficient and popular method. Usually, there are two technical routes for the construction of defect detection algorithm models: one is to use a lightweight network, which improves the efficiency, but it can generally only target a few types of defects and may reduce the detection accuracy; the other is to use a complex network model, which improves the accuracy, and can identify multiple types of defects at the same time, but it has a large computational volume and low efficiency. To maintain the model's high detection accuracy as well as its lightweight structure, this paper proposes a lightweight and efficient multi type defect detection method for transmission lines based on DCP-YOLOv8. The method employs deformable convolution (C2f_DCNv3) to enhance the defect feature extraction capability, and designs a re-parameterized cross phase feature fusion structure (RCSP) to optimize and fuse high-level semantic features with low level spatial features, thus improving the capability of the model to recognize defects at different scales while significantly reducing the model parameters; additionally, it combines the dynamic detection head and deformable convolutional v3's detection head (DCNv3-Dyhead) to enhance the feature expression capability and the utilization of contextual information to further improve the detection accuracy. Experimental results show that on a dataset containing 20 real transmission line defects, the method increases the average accuracy (mAP@0.5) to 72.2%, an increase of 4.3%, compared with the lightest baseline YOLOv8n model; the number of model parameters is only 2.8 M, a reduction of 9.15%, and the number of processed frames per second (FPS) reaches 103, which meets the real time detection demand. In the scenario of multi type defect detection, it effectively balances detection accuracy and performance with quantitative generalizability. [ABSTRACT FROM AUTHOR]

The correct workability of concrete is an essential parameter for its placement and compaction. However, an absence of automatic and transparent measurement methods to estimate the workability of concrete hinders the adaptation from laborious traditional methods such as the slump test. In this paper, we developed a machine-learning framework for estimating the slump class of concrete in the mixer using a stereovision camera. Depth data from five different slump classes was transformed into Haralick texture features to train several machine-learning classifiers. The best-performing classifier achieved a multiclass classification accuracy of 0.8179 with the XGBoost algorithm. Furthermore, we found through statistical analysis that while the denoising of depth data has little effect on the accuracy, the feature extraction of mixer blades and the choice of region of interest significantly increase the accuracy and the efficiency of the classifiers. The proposed framework shows robust results, indicating that stereovision is a competitive solution to estimate the workability of concrete during concrete production. [ABSTRACT FROM AUTHOR]

Human–object interaction (HOI) detection identifies a "set of interactions" in an image involving the recognition of interacting instances and the classification of interaction categories. The complexity and variety of image content make this task challenging. Recently, the Transformer has been applied in computer vision and received attention in the HOI detection task. Therefore, this paper proposes a novel Part Refinement Tandem Transformer (PRTT) for HOI detection. Unlike the previous Transformer-based HOI method, PRTT utilizes multiple decoders to split and process rich elements of HOI prediction and introduces a new part state feature extraction (PSFE) module to help improve the final interaction category classification. We adopt a novel prior feature integrated cross-attention (PFIC) to utilize the fine-grained partial state semantic and appearance feature output obtained by the PSFE module to guide queries. We validate our method on two public datasets, V-COCO and HICO-DET. Compared to state-of-the-art models, the performance of detecting human–object interaction is significantly improved by the PRTT. [ABSTRACT FROM AUTHOR]

Crowd counting, as a basic computer vision task, plays an important role in many fields such as video surveillance, accident prediction, public security, and intelligent transportation. At present, crowd counting tasks face various challenges. Firstly, due to the diversity of crowd distribution and increasing population density, there is a phenomenon of large-scale crowd aggregation in public places, sports stadiums, and stations, resulting in very serious occlusion. Secondly, when annotating large-scale datasets, positioning errors can also easily affect training results. In addition, the size of human head targets in dense images is not consistent, making it difficult to identify both near and far targets using only one network simultaneously. The existing crowd counting methods mainly use density plot regression methods. However, this framework does not distinguish the features between distant and near targets and cannot adaptively respond to scale changes. Therefore, the detection performance in areas with sparse population distribution is not good. To solve such problems, we propose an adaptive multi-scale far and near distance network based on the convolutional neural network (CNN) framework for counting dense populations and achieving a good balance between accuracy, inference speed, and performance. However, on the feature level, in order to enable the model to distinguish the differences between near and far features, we use stacked convolution layers to deepen the depth of the network, allocate different receptive fields according to the distance between the target and the camera, and fuse the features between nearby targets to enhance the feature extraction ability of pedestrians under nearby targets. Secondly, depth information is used to distinguish distant and near targets of different scales and the original image is cut into four different patches to perform pixel-level adaptive modeling on the population. In addition, we add density normalized average precision (nAP) indicators to analyze the accuracy of our method in spatial positioning. This paper validates the effectiveness of NF-Net on three challenging benchmarks in Shanghai Tech Part A and B, UCF_ CC_50, and UCF-QNRF datasets. Compared with SOTA, it has more significant performance in various scenarios. In the UCF-QNRF dataset, it is further validated that our method effectively solves the interference of complex backgrounds. [ABSTRACT FROM AUTHOR]

Blurring is one of the main degradation factors in image degradation, so image deblurring is of great interest as a fundamental problem in low-level computer vision. Because of the limited receptive field, traditional CNNs lack global fuzzy region modeling, and do not make full use of rich context information between features. Recently, a transformer-based neural network structure has performed well in natural language tasks, inspiring rapid development in the field of defuzzification. Therefore, in this paper, a hybrid architecture based on CNN and transformers is used for image deblurring. Specifically, we first extract the shallow features of the blurred images using a cross-layer feature fusion block that emphasizes the contextual information of each feature extraction layer. Secondly, an efficient transformer module for extracting deep features is designed, which fully aggregates feature information at medium and long distances using vertical and horizontal intra- and inter-strip attention layers, and a dual gating mechanism is used as a feedforward neural network, which effectively reduces redundant features. Finally, the cross-layer feature fusion block is used to complement the feature information to obtain the deblurred image. Extensive experimental results on publicly available benchmark datasets GoPro, HIDE, and the real dataset RealBlur show that the proposed method outperforms the current mainstream deblurring algorithms and recovers the edge contours and texture details of the images more clearly. [ABSTRACT FROM AUTHOR]

Image recognition has long been one of the research hotspots in computer vision tasks. The development of deep learning is rapid in recent years, and convolutional neural networks usually need to be designed with fixed resources. If sufficient resources are available, the model can be scaled up to achieve higher accuracy, for example, VggNet, ResNet, GoogLeNet, etc. Although the accuracy of large-scale models has been improved, the following problems will occur with the expansion of model scale: (1) There may be over-fitting; (2) increasing model parameters; (3) slow model convergence. This paper proposes a design method for a modular convolutional neural network model which solves the problem of over-fitting and large model parameters by connecting multiple modules in parallel. Moreover, each module contains several submodules (three submodules in this paper) and fuses the features extracted from the submodules. The model convergence can be accelerated by using the fused features (the fused features contain more image information). In this study, we add a gate unit based on the attention mechanism to the model, which aims to optimize the structure of the model (select the optimal number of modules), allowing the model to select an optimum network structure by learning and dynamically reducing FLOPs (floating-point operations per second) of the model. Compared to VggNet, ResNet, and GoogLeNet, the structure of the model proposed in this paper is simple and the parameters are small. The proposed model achieves good results in the Kaggle datasets Cats-vs.-Dogs (99.3%), 10-Monkey Species (99.26%), and Birds-400 (99.13%). [ABSTRACT FROM AUTHOR]

Instance segmentation is a challenging task in computer vision, as it requires distinguishing objects and predicting dense areas. Currently, segmentation models based on complex designs and large parameters have achieved remarkable accuracy. However, from a practical standpoint, achieving a balance between accuracy and speed is even more desirable. To address this need, this paper presents ESAMask, a real-time segmentation model fused with efficient sparse attention, which adheres to the principles of lightweight design and efficiency. In this work, we propose several key contributions. Firstly, we introduce a dynamic and sparse Related Semantic Perceived Attention mechanism (RSPA) for adaptive perception of different semantic information of various targets during feature extraction. RSPA uses the adjacency matrix to search for regions with high semantic correlation of the same target, which reduces computational cost. Additionally, we design the GSInvSAM structure to reduce redundant calculations of spliced features while enhancing interaction between channels when merging feature layers of different scales. Lastly, we introduce the Mixed Receptive Field Context Perception Module (MRFCPM) in the prototype branch to enable targets of different scales to capture the feature representation of the corresponding area during mask generation. MRFCPM fuses information from three branches of global content awareness, large kernel region awareness, and convolutional channel attention to explicitly model features at different scales. Through extensive experimental evaluation, ESAMask achieves a mask AP of 45.4 at a frame rate of 45.2 FPS on the COCO dataset, surpassing current instance segmentation methods in terms of the accuracy–speed trade-off, as demonstrated by our comprehensive experimental results. In addition, the high-quality segmentation results of our proposed method for objects of various classes and scales can be intuitively observed from the visualized segmentation outputs. [ABSTRACT FROM AUTHOR]

Autonomous vehicles have become a topic of interest in recent times due to the rapid advancement of automobile and computer vision technology. The ability of autonomous vehicles to drive safely and efficiently relies heavily on their ability to accurately recognize traffic signs. This makes traffic sign recognition a critical component of autonomous driving systems. To address this challenge, researchers have been exploring various approaches to traffic sign recognition, including machine learning and deep learning. Despite these efforts, the variability of traffic signs across different geographical regions, complex background scenes, and changes in illumination still poses significant challenges to the development of reliable traffic sign recognition systems. This paper provides a comprehensive overview of the latest advancements in the field of traffic sign recognition, covering various key areas, including preprocessing techniques, feature extraction methods, classification techniques, datasets, and performance evaluation. The paper also delves into the commonly used traffic sign recognition datasets and their associated challenges. Additionally, this paper sheds light on the limitations and future research prospects of traffic sign recognition. [ABSTRACT FROM AUTHOR]

Due to the sharp increase in household waste, its separate collection is essential in order to reduce the huge amount of household waste, since it is difficult to recycle trash without separate collection. However, since it is costly and time-consuming to separate trash manually, it is crucial to develop an automatic system for separate collection using deep learning and computer vision. In this paper, we propose two Anchor-free-based Recyclable Trash Detection Networks (ARTD-Net) which can recognize overlapped multiple wastes of different types efficiently by using edgeless modules: ARTD-Net1 and ARTD-Net2. The former is an anchor-free based one-stage deep learning model which consists of three modules: centralized feature extraction, multiscale feature extraction and prediction. The centralized feature extraction module in backbone architecture focuses on extracting features around the center of the input image to improve detection accuracy. The multiscale feature extraction module provides feature maps of different scales through bottom-up and top-down pathways. The prediction module improves classification accuracy of multiple objects based on edge weights adjustments for each instance. The latter is an anchor-free based multi-stage deep learning model which can efficiently finds each of waste regions by additionally exploiting region proposal network and RoIAlign. It sequentially performs classification and regression to improve accuracy. Therefore, ARTD-Net2 is more accurate than ARTD-Net1, while ARTD-Net1 is faster than ARTD-Net2. We shall show that our proposed ARTD-Net1 and ARTD-Net2 methods achieve competitive performance in mean average precision and F1 score compared to other deep learning models. The existing datasets have several problems that do not deal with the important class of wastes produced commonly in the real world, and they also do not consider the complex arrangement of multiple wastes with different types. Moreover, most of the existing datasets have an insufficient number of images with low resolution. We shall present a new recyclables dataset which is composed of a large number of high-resolution waste images with additional essential classes. We shall show that waste detection performance is improved by providing various images with the complex arrangement of overlapped multiple wastes with different types. [ABSTRACT FROM AUTHOR]

Computer vision technology is increasingly being used in areas such as intelligent security and autonomous driving. Users need accurate and reliable visual information, but the images obtained under severe weather conditions are often disturbed by rainy weather, causing image scenes to look blurry. Many current single image deraining algorithms achieve good performance but have limitations in retaining detailed image information. In this paper, we design a Scale-space Feature Recalibration Network (SFR-Net) for single image deraining. The proposed network improves the image feature extraction and characterization capability of a Multi-scale Extraction Recalibration Block (MERB) using dilated convolution with different convolution kernel sizes, which results in rich multi-scale rain streaks features. In addition, we develop a Subspace Coordinated Attention Mechanism (SCAM) and embed it into MERB, which combines coordinated attention recalibration and a subspace attention mechanism to recalibrate the rain streaks feature information learned from the feature extraction phase and eliminate redundant feature information to enhance the transfer of important feature information. Meanwhile, the overall SFR-Net structure uses dense connection and cross-layer feature fusion to repeatedly utilize the feature maps, thus enhancing the understanding of the network and avoiding gradient disappearance. Through extensive experiments on synthetic and real datasets, the proposed method outperforms the recent state-of-the-art deraining algorithms in terms of both the rain removal effect and the preservation of image detail information. [ABSTRACT FROM AUTHOR]

Unsupervised domain adaptation (UDA) aims to mitigate the performance drop due to the distribution shift between the training and testing datasets. UDA methods have achieved performance gains for models trained on a source domain with labeled data to a target domain with only unlabeled data. The standard feature extraction method in domain adaptation has been convolutional neural networks (CNNs). Recently, attention-based transformer models have emerged as effective alternatives for computer vision tasks. In this paper, we benchmark three attention-based architectures, specifically vision transformer (ViT), shifted window transformer (SWIN), and dual attention vision transformer (DAViT), against convolutional architectures ResNet, HRNet and attention-based ConvNext, to assess the performance of different backbones for domain generalization and adaptation. We incorporate these backbone architectures as feature extractors in the source hypothesis transfer (SHOT) framework for UDA. SHOT leverages the knowledge learned in the source domain to align the image features of unlabeled target data in the absence of source domain data, using self-supervised deep feature clustering and self-training. We analyze the generalization and adaptation performance of these models on standard UDA datasets and aerial UDA datasets. In addition, we modernize the training procedure commonly seen in UDA tasks by adding image augmentation techniques to help models generate richer features. Our results show that ConvNext and SWIN offer the best performance, indicating that the attention mechanism is very beneficial for domain generalization and adaptation with both transformer and convolutional architectures. Our ablation study shows that our modernized training recipe, within the SHOT framework, significantly boosts performance on aerial datasets. [ABSTRACT FROM AUTHOR]

Three-dimensional face recognition is an important part of the field of computer vision. Point clouds are widely used in the field of 3D vision due to the simple mathematical expression. However, the disorder of the points makes it difficult for them to have ordered indexes in convolutional neural networks. In addition, the point clouds lack detailed textures, which makes the facial features easily affected by expression or head pose changes. To solve the above problems, this paper constructs a new face recognition network, which mainly consists of two parts. The first part is a novel operator based on a local feature descriptor to realize the fine-grained features extraction and the permutation invariance of point clouds. The second part is a feature enhancement mechanism to enhance the discrimination of facial features. In order to verify the performance of our method, we conducted experiments on three public datasets: CASIA-3D, Bosphorus, and Lock3Dface. The results show that the accuracy of our method is improved by 0.7%, 0.4%, and 0.8% compared with the latest methods on these three datasets, respectively. [ABSTRACT FROM AUTHOR]

Operating in extreme environments is often challenging due to the lack of perceptual knowledge. During fire incidents in large buildings, the extreme levels of smoke can seriously impede a firefighter's vision, potentially leading to severe material damage and loss of life. To increase the safety of firefighters, research is conducted in collaboration with Dutch fire departments into the usability of Unmanned Ground Vehicles to increase situational awareness in hazardous environments. This paper proposes FirebotSLAM, the first algorithm capable of coherently computing a robot's odometry while creating a comprehensible 3D map solely using the information extracted from thermal images. The literature showed that the most challenging aspect of thermal Simultaneous Localization and Mapping (SLAM) is the extraction of robust features in thermal images. Therefore, a practical benchmark of feature extraction and description methods was performed on datasets recorded during a fire incident. The best-performing combination of extractor and descriptor is then implemented into a state-of-the-art visual SLAM algorithm. As a result, FirebotSLAM is the first thermal odometry algorithm able to perform global trajectory optimization by detecting loop closures. Finally, FirebotSLAM is the first thermal SLAM algorithm to be tested in a fiery environment to validate its applicability in an operational scenario. [ABSTRACT FROM AUTHOR]

Traffic sign detection is a crucial task in computer vision, finding wide-ranging applications in intelligent transportation systems, autonomous driving, and traffic safety. However, due to the complexity and variability of traffic environments and the small size of traffic signs, detecting small traffic signs in real-world scenes remains a challenging problem. In order to improve the recognition of road traffic signs, this paper proposes a small object detection algorithm for traffic signs based on the improved YOLOv7. First, the small target detection layer in the neck region was added to augment the detection capability for small traffic sign targets. Simultaneously, the integration of self-attention and convolutional mix modules (ACmix) was applied to the newly added small target detection layer, enabling the capture of additional feature information through the convolutional and self-attention channels within ACmix. Furthermore, the feature extraction capability of the convolution modules was enhanced by replacing the regular convolution modules in the neck layer with omni-dimensional dynamic convolution (ODConv). To further enhance the accuracy of small target detection, the normalized Gaussian Wasserstein distance (NWD) metric was introduced to mitigate the sensitivity to minor positional deviations of small objects. The experimental results on the challenging public dataset TT100K demonstrate that the SANO-YOLOv7 algorithm achieved an 88.7% mAP@0.5, outperforming the baseline model YOLOv7 by 5.3%. [ABSTRACT FROM AUTHOR]

Scene text recognition is a crucial area of research in computer vision. However, current mainstream scene text recognition models suffer from incomplete feature extraction due to the small downsampling scale used to extract features and obtain more features. This limitation hampers their ability to extract complete features of each character in the image, resulting in lower accuracy in the text recognition process. To address this issue, a novel text recognition model based on multi-scale fusion and the convolutional recurrent neural network (CRNN) has been proposed in this paper. The proposed model has a convolutional layer, a feature fusion layer, a recurrent layer, and a transcription layer. The convolutional layer uses two scales of feature extraction, which enables it to derive two distinct outputs for the input text image. The feature fusion layer fuses the different scales of features and forms a new feature. The recurrent layer learns contextual features from the input sequence of features. The transcription layer outputs the final result. The proposed model not only expands the recognition field but also learns more image features at different scales; thus, it extracts a more complete set of features and achieving better recognition of text. The results of experiments are then presented to demonstrate that the proposed model outperforms the CRNN model on text datasets, such as Street View Text, IIIT-5K, ICDAR2003, and ICDAR2013 scenes, in terms of text recognition accuracy. [ABSTRACT FROM AUTHOR]

The prices of different quality river crabs on the market can vary several times. Therefore, the internal quality identification and accurate sorting of crabs are particularly important for improving the economic benefits of the industry. Using existing sorting methods by labor and weight to meet the urgent needs of mechanization and intelligence in the crab breeding industry is difficult. Therefore, this paper proposes an improved BP neural network model based on a genetic algorithm, which can grade the crab quality. We comprehensively considered the four characteristics of crabs as the input variables of the model, namely gender, fatness, weight, and shell color of crabs, among which gender, fatness, and shell color were obtained by image processing technology, whereas weight is obtained using a load cell. First, mature machine vision technology is used to preprocess the images of the crab's abdomen and back, and then feature information is extracted from the images. Next, genetic and backpropagation algorithms are combined to establish a quality grading model for crab, and data training is conducted on the model to obtain the optimal threshold and weight values. Analysis of experimental results reveals that the average classification accuracy reaches 92.7%, which proves that this method can achieve efficient and accurate classification and sorting of crabs, successfully addressing market demand. [ABSTRACT FROM AUTHOR]

The counting of surgical instruments is an important task to ensure surgical safety and patient health. However, due to the uncertainty of manual operations, there is a risk of missing or miscounting instruments. Applying computer vision technology to the instrument counting process can not only improve efficiency, but also reduce medical disputes and promote the development of medical informatization. However, during the counting process, surgical instruments may be densely arranged or obstruct each other, and they may be affected by different lighting environments, all of which can affect the accuracy of instrument recognition. In addition, similar instruments may have only minor differences in appearance and shape, which increases the difficulty of identification. To address these issues, this paper improves the YOLOv7x object detection algorithm and applies it to the surgical instrument detection task. First, the RepLK Block module is introduced into the YOLOv7x backbone network, which can increase the effective receptive field and guide the network to learn more shape features. Second, the ODConv structure is introduced into the neck module of the network, which can significantly enhance the feature extraction ability of the basic convolution operation of the CNN and capture more rich contextual information. At the same time, we created the OSI26 data set, which contains 452 images and 26 surgical instruments, for model training and evaluation. The experimental results show that our improved algorithm exhibits higher accuracy and robustness in surgical instrument detection tasks, with F1, AP, AP50, and AP75 reaching 94.7%, 91.5%, 99.1%, and 98.2%, respectively, which are 4.6%, 3.1%, 3.6%, and 3.9% higher than the baseline. Compared to other mainstream object detection algorithms, our method has significant advantages. These results demonstrate that our method can more accurately identify surgical instruments, thereby improving surgical safety and patient health. [ABSTRACT FROM AUTHOR]

Industrial inspection is crucial for maintaining quality and safety in industrial processes. Deep learning models have recently demonstrated promising results in such tasks. This paper proposes YOLOX-Ray, an efficient new deep learning architecture tailored for industrial inspection. YOLOX-Ray is based on the You Only Look Once (YOLO) object detection algorithms and integrates the SimAM attention mechanism for improved feature extraction in the Feature Pyramid Network (FPN) and Path Aggregation Network (PAN). Moreover, it also employs the Alpha-IoU cost function for enhanced small-scale object detection. YOLOX-Ray's performance was assessed in three case studies: hotspot detection, infrastructure crack detection and corrosion detection. The architecture outperforms all other configurations, achieving m A P 50 values of 89%, 99.6% and 87.7%, respectively. For the most challenging metric, m A P 50 : 95 , the achieved values were 44.7%, 66.1% and 51.8%, respectively. A comparative analysis demonstrated the importance of combining the SimAM attention mechanism with Alpha-IoU loss function for optimal performance. In conclusion, YOLOX-Ray's ability to detect and to locate multi-scale objects in industrial environments presents new opportunities for effective, efficient and sustainable inspection processes across various industries, revolutionizing the field of industrial inspections. [ABSTRACT FROM AUTHOR]

It is considered that 1 in 10 adults worldwide have diabetes. Diabetic foot ulcers are some of the most common complications of diabetes, and they are associated with a high risk of lower-limb amputation and, as a result, reduced life expectancy. Timely detection and periodic ulcer monitoring can considerably decrease amputation rates. Recent research has demonstrated that computer vision can be used to identify foot ulcers and perform non-contact telemetry by using ulcer and tissue area segmentation. However, the applications are limited to controlled lighting conditions, and expert knowledge is required for dataset annotation. This paper reviews the latest publications on the use of artificial intelligence for ulcer area detection and segmentation. The PRISMA methodology was used to search for and select articles, and the selected articles were reviewed to collect quantitative and qualitative data. Qualitative data were used to describe the methodologies used in individual studies, while quantitative data were used for generalization in terms of dataset preparation and feature extraction. Publicly available datasets were accounted for, and methods for preprocessing, augmentation, and feature extraction were evaluated. It was concluded that public datasets can be used to form a bigger, more diverse datasets, and the prospects of wider image preprocessing and the adoption of augmentation require further research. [ABSTRACT FROM AUTHOR]

Recently, the scientific community has placed great emphasis on the recognition of human activity, especially in the area of health and care for the elderly. There are already practical applications of activity recognition and unusual conditions that use body sensors such as wrist-worn devices or neck pendants. These relatively simple devices may be prone to errors, might be uncomfortable to wear, might be forgotten or not worn, and are unable to detect more subtle conditions such as incorrect postures. Therefore, other proposed methods are based on the use of images and videos to carry out human activity recognition, even in open spaces and with multiple people. However, the resulting increase in the size and complexity involved when using image data requires the use of the most recent advanced machine learning and deep learning techniques. This paper presents an approach based on deep learning with attention to the recognition of activities from multiple frames. Feature extraction is performed by estimating the pose of the human skeleton, and classification is performed using a neural network based on Bidirectional Encoder Representation of Transformers (BERT). This algorithm was trained with the UP-Fall public dataset, generating more balanced artificial data with a Generative Adversarial Neural network (GAN), and evaluated with real data, outperforming the results of other activity recognition methods using the same dataset. [ABSTRACT FROM AUTHOR]

With a large number of Local Binary Patterns (LBP) variants being currently used today, the significant and importance of visual descriptors in computer vision applications are prominent. This paper presents a novel visual descriptor, i.e., SIM-LBP. It employs a new matrix technique called the Symmetric Inline Matrix generator method, which acts as a new variant of LBP. The key feature that separates our variant from existing counterparts is that our variant is very efficient in extracting facial expression features like eyes, eye brows, nose and mouth in a wide range of lighting conditions. For testing our model, we applied SIM-LBP on the JAFFE dataset to convert all the images to its corresponding SIM-LBP transformed variant. These transformed images are then used to train a Convolution Neural Network (CNN) based deep learning model for facial expressions recognition (FER). Several performance evaluation metrics, i.e., recognition accuracy rate, precision, recall, and F1-score, were used to test mode efficiency in comparison with those using the traditional LBP descriptor and other LBP variants. Our model outperformed in all four matrices with the proposed SIM-LBP transformation on the input images against those of baseline methods. In comparison analysis with the other state-of-the-art methods, it shows the usefulness of the proposed SIM-LBP model. Our proposed SIM-LBP variant transformation can also be applied on facial images to identify a person's mental states and predict mood variations. [ABSTRACT FROM AUTHOR]

This paper discusses an active droplet generation system, and the presented droplet generator successfully performs droplet generation using two fluid phases: continuous phase fluid and dispersed phase fluid. The performance of an active droplet generation system is analysed based on the droplet morphology using vision sensing and digital image processing. The proposed system in the study includes a droplet generator, camera module with image pre-processing and identification algorithm, and controller and control algorithm with a workstation computer. The overall system is able to control, sense, and analyse the generation of droplets. The main controller consists of a microcontroller, motor controller, voltage regulator, and power supply. Among the morphological features of droplets, the diameter is extracted from the images to observe the system performance. The MATLAB-based image processing algorithm consists of image acquisition, image enhancement, droplet identification, feature extraction, and analysis. RGB band filtering, thresholding, and opening are used in image pre-processing. After the image enhancement, droplet identification is performed by tracing the boundary of the droplets. The average droplet diameter varied from ~3.05 mm to ~4.04 mm in the experiments, and the average droplet diameter decrement presented a relationship of a second-order polynomial with the droplet generation time. [ABSTRACT FROM AUTHOR]

In modern cyber-physical systems, the integration of AI into vision pipelines is now a standard practice for applications ranging from autonomous vehicles to mobile devices. Traditional AI integration often relies on cloud-based processing, which faces challenges such as data access bottlenecks, increased latency, and high power consumption. This article reviews embedded AI vision systems, examining the diverse landscape of near-sensor and in-sensor processing architectures that incorporate convolutional neural networks. We begin with a comprehensive analysis of the critical characteristics and metrics that define the performance of AI-integrated vision systems. These include sensor resolution, frame rate, data bandwidth, computational throughput, latency, power efficiency, and overall system scalability. Understanding these metrics provides a foundation for evaluating how different embedded processing architectures impact the entire vision pipeline, from image capture to AI inference. Our analysis delves into near-sensor systems that leverage dedicated hardware accelerators and commercially available components to efficiently process data close to their source, minimizing data transfer overhead and latency. These systems offer a balance between flexibility and performance, allowing for real-time processing in constrained environments. In addition, we explore in-sensor processing solutions that integrate computational capabilities directly into the sensor. This approach addresses the rigorous demand constraints of embedded applications by significantly reducing data movement and power consumption while also enabling in-sensor feature extraction, pre-processing, and CNN inference. By comparing these approaches, we identify trade-offs related to flexibility, power consumption, and computational performance. Ultimately, this article provides insights into the evolving landscape of embedded AI vision systems and suggests new research directions for the development of next-generation machine vision systems. [ABSTRACT FROM AUTHOR]

This paper addresses the problem of automatic quality inspection in assembly processes by discussing the design of a computer vision system realized by means of a heterogeneous multiprocessor system-on-chip. Such an approach was applied to a real catalytic converter assembly process, to detect planar, translational, and rotational shifts of the flanges welded on the central body. The manufacturing line imposed tight time and room constraints. The image processing method and the features extraction algorithm, based on a specific geometrical model, are described and validated. The algorithm was developed to be highly modular, thus suitable to be implemented by adopting a hardware–software co-design strategy. The most timing consuming computational steps were identified and then implemented by dedicated hardware accelerators. The entire system was implemented on a Xilinx Zynq heterogeneous system-on-chip by using a hardware–software (HW–SW) co-design approach. The system is able to detect planar and rotational shifts of welded flanges, with respect to the ideal positions, with a maximum error lower than one millimeter and one sexagesimal degree, respectively. Remarkably, the proposed HW–SW approach achieves a 23× speed-up compared to the pure software solution running on the Zynq embedded processing system. Therefore, it allows an in-line automatic quality inspection to be performed without affecting the production time of the existing manufacturing process. [ABSTRACT FROM AUTHOR]

Transformer-based methodologies in object detection have recently piqued considerable interest and have produced impressive results. DETR, an end-to-end object detection framework, ingeniously integrates the Transformer architecture, traditionally used in NLP, into computer vision for sequence-to-sequence prediction. Its enhanced variant, DINO, featuring improved denoising anchor boxes, has showcased remarkable performance on the COCO val2017 dataset. However, it often encounters challenges when applied to scenarios involving small object detection. Thus, we propose an innovative method for feature enhancement tailored to recursive prediction tasks, with a particular emphasis on augmenting small object detection performance. It primarily involves three enhancements: refining the backbone to favor feature maps that are more sensitive to small targets, incrementally augmenting the number of queries for small objects, and advancing the loss function for better performance. Specifically, The study incorporated the Switchable Atrous Convolution (SAC) mechanism, which features adaptable dilated convolutions, to increment the receptive field and thus elevate the innate feature extraction capabilities of the primary network concerning diminutive objects. Subsequently, a Recursive Small Object Prediction (RSP) module was designed to enhance the feature extraction of the prediction head for more precise network operations. Finally, the loss function was augmented with the Normalized Wasserstein Distance (NWD) metric, tailoring the loss function to suit small object detection better. The efficacy of the proposed model is empirically confirmed via testing on the VISDRONE2019 dataset. The comprehensive array of experiments indicates that our proposed model outperforms the extant DINO model in terms of average precision (AP) small object detection. [ABSTRACT FROM AUTHOR]

Video action recognition based on skeleton nodes is a highlighted issue in the computer vision field. In real application scenarios, the large number of skeleton nodes and behavior occlusion problems between individuals seriously affect recognition speed and accuracy. Therefore, we proposed a lightweight multi-stream feature cross-fusion (L-MSFCF) model to recognize abnormal behaviors such as fighting, vicious kicking, climbing over the wall, et al., which could obviously improve recognition speed based on lightweight skeleton node calculation, and improve recognition accuracy based on occluded skeleton node prediction analysis in order to effectively solve the behavior occlusion problem. The experiments show that our proposed All-MSFCF model has a video action recognition average accuracy rate of 92.7% for eight kinds of abnormal behavior recognition. Although our proposed lightweight L-MSFCF model has an 87.3% average accuracy rate, its average recognition speed is 62.7% higher than the full-skeleton recognition model, which is more suitable for solving real-time tracing problems. Moreover, our proposed Trajectory Prediction Tracking (TPT) model could real-time predict the moving positions based on the dynamically selected core skeleton node calculation, especially for the short-term prediction within 15 frames and 30 frames that have lower average loss errors. [ABSTRACT FROM AUTHOR]

Point cloud registration is an important task in computer vision and robotics which is widely used in 3D reconstruction, target recognition, and other fields. At present, many registration methods based on deep learning have better registration accuracy in complete point cloud registration, but partial registration accuracy is poor. Therefore, a partial point cloud registration network, HALNet, is proposed. Firstly, a feature extraction network consisting mainly of adaptive graph convolution (AGConv), two-dimensional convolution, and convolution block attention (CBAM) is used to learn the features of the initial point cloud. Then the overlapping estimation is used to remove the non-overlapping points of the two point clouds, and the hybrid attention mechanism composed of self-attention and cross-attention is used to fuse the geometric information of the two point clouds. Finally, the rigid transformation is obtained by using the fully connected layer. Five methods with excellent registration performance were selected for comparison. Compared with SCANet, which has the best registration performance among the five methods, the RMSE(R) and MAE(R) of HALNet are reduced by 10.67% and 12.05%. In addition, the results of the ablation experiment verify that the hybrid attention mechanism and fully connected layer are conducive to improving registration performance. [ABSTRACT FROM AUTHOR]

Superpixel segmentation is one of the key image preprocessing steps in object recognition and detection methods. However, the over-segmentation in the smoothly connected homogenous region in an image is the key problem. That would produce redundant complex jagged textures. In this paper, the density peak clustering will be used to reduce the redundant superpixels and highlight the primary textures and contours of the salient objects. Firstly, the grid pixels are extracted as feature points, and the density of each feature point will be defined. Secondly, the cluster centers are extracted with the density peaks. Finally, all the feature points will be clustered by the density peaks. The pixel blocks, which are obtained by the above steps, are superpixels. The method is carried out in the BSDS500 dataset, and the experimental results show that the Boundary Recall (BR) and Achievement Segmentation Accuracy (ASA) are 95.0% and 96.3%, respectively. In addition, the proposed method has better performance in efficiency (30 fps). The comparison experiments show that not only do the superpixel boundaries have good adhesion to the primary textures and contours of the salient objects, but they can also effectively reduce the redundant superpixels in the homogeneous region. [ABSTRACT FROM AUTHOR]

Sign language recognition, an essential interface between the hearing and deaf-mute communities, faces challenges with high false positive rates and computational costs, even with the use of advanced deep learning techniques. Our proposed solution is a stacked encoded model, combining artificial intelligence (AI) with the Internet of Things (IoT), which refines feature extraction and classification to overcome these challenges. We leverage a lightweight backbone model for preliminary feature extraction and use stacked autoencoders to further refine these features. Our approach harnesses the scalability of big data, showing notable improvement in accuracy, precision, recall, F1-score, and complexity analysis. Our model's effectiveness is demonstrated through testing on the ArSL2018 benchmark dataset, showcasing superior performance compared to state-of-the-art approaches. Additional validation through an ablation study with pre-trained convolutional neural network (CNN) models affirms our model's efficacy across all evaluation metrics. Our work paves the way for the sustainable development of high-performing, IoT-based sign-language-recognition applications. [ABSTRACT FROM AUTHOR]

For facing of the problems caused by the YOLOv4 algorithm's insensitivity to small objects and low detection precision in traffic light detection and recognition, the Improved YOLOv4 algorithm is investigated in the paper using the shallow feature enhancement mechanism and the bounding box uncertainty prediction mechanism. The shallow feature enhancement mechanism is used to extract features from the network and improve the network's ability to locate small objects and color resolution by merging two shallow features at different stages with the high-level semantic features obtained after two rounds of upsampling. Uncertainty is introduced in the bounding box prediction mechanism to improve the reliability of the prediction of the bounding box by modeling the output coordinates of the prediction bounding box and adding the Gaussian model to calculate the uncertainty of the coordinate information. The LISA traffic light data set is used to perform detection and recognition experiments separately. The Improved YOLOv4 algorithm is shown to have a high effectiveness in enhancing the detection and recognition precision of traffic lights. In the detection experiment, the area under the PR curve value of the Improved YOLOv4 algorithm is found to be 97.58%, which represents an increase of 7.09% in comparison to the 90.49% score gained in the Vision for Intelligent Vehicles and Applications Challenge Competition. In the recognition experiment, the mean average precision of the Improved YOLOv4 algorithm is 82.15%, which is 2.86% higher than that of the original YOLOv4 algorithm. The Improved YOLOv4 algorithm shows remarkable advantages as a robust and practical method for use in the real-time detection and recognition of traffic signal lights. [ABSTRACT FROM AUTHOR]

The existing classification methods for Panax notoginseng taproots suffer from low accuracy, low efficiency, and poor stability. In this study, a classification model based on image feature fusion is established for Panax notoginseng taproots. The images of Panax notoginseng taproots collected in the experiment are preprocessed by Gaussian filtering, binarization, and morphological methods. Then, a total of 40 features are extracted, including size and shape features, HSV and RGB color features, and texture features. Through BP neural network, extreme learning machine (ELM), and support vector machine (SVM) models, the importance of color, texture, and fusion features for the classification of the main roots of Panax notoginseng is verified. Among the three models, the SVM model performs the best, achieving an accuracy of 92.037% on the prediction set. Next, iterative retaining information variables (IRIVs), variable iterative space shrinkage approach (VISSA), and stepwise regression analysis (SRA) are used to reduce the dimension of all the features. Finally, a traditional machine learning SVM model based on feature selection and a deep learning model based on semantic segmentation are established. With the model size of only 125 kb and the training time of 3.4 s, the IRIV-SVM model achieves an accuracy of 95.370% on the test set, so IRIV-SVM is selected as the main root classification model for Panax notoginseng. After being optimized by the gray wolf optimizer, the IRIV-GWO-SVM model achieves the highest classification accuracy of 98.704% on the test set. The study results of this paper provide a basis for developing online classification methods of Panax notoginseng with different grades in actual production. [ABSTRACT FROM AUTHOR]

Reliable tools for artefact rejection and signal classification are a must for cosmic ray detection experiments based on CMOS technology. In this paper, we analyse the fitness of several feature-based statistical classifiers for the classification of particle candidate hits in four categories: spots, tracks, worms and artefacts. We use Zernike moments of the image function as feature carriers and propose a preprocessing and denoising scheme to make the feature extraction more efficient. As opposed to convolution neural network classifiers, the feature-based classifiers allow for establishing a connection between features and geometrical properties of candidate hits. Apart from basic classifiers we also consider their ensemble extensions and find these extensions generally better performing than basic versions, with an average recognition accuracy of 88%. [ABSTRACT FROM AUTHOR]

Due to its advantages of low latency, low power consumption, and high flexibility, FPGA-based acceleration technology has been more and more widely studied and applied in the field of computer vision in recent years. An FPGA-based feature extraction and tracking accelerator for real-time visual odometry (VO) and visual simultaneous localization and mapping (V-SLAM) is proposed, which can realize the complete acceleration processing capability of the image front-end. For the first time, we implement a hardware solution that combines features from accelerated segment test (FAST) feature points with Gunnar Farneback (GF) dense optical flow to achieve better feature tracking performance and provide more flexible technical route selection. In order to solve the scale invariance and rotation invariance lacking problems of FAST features, an efficient pyramid module with a five-layer thumbnail structure was designed and implemented. The accelerator was implemented on a modern Xilinx Zynq FPGA. The evaluation results showed that the accelerator could achieve stable tracking of features of violently shaking images and were consistent with the results from MATLAB code running on PCs. Compared to PC CPUs, which require seconds of processing time, the processing latency was greatly reduced to the order of milliseconds, making GF dense optical flow an efficient and practical technical solution on the edge side. [ABSTRACT FROM AUTHOR]

Video analytics and computer vision applications face challenges when using video sequences with low visibility. The visibility of a video sequence is degraded when the sequence is affected by atmospheric interference like rain. Many approaches have been proposed to remove rain streaks from video sequences. Some approaches are based on physical features, and some are based on data-driven (i.e., deep-learning) models. Although the physical features-based approaches have better rain interpretability, the challenges are extracting the appropriate features and fusing them for meaningful rain removal, as the rain streaks and moving objects have dynamic physical characteristics and are difficult to distinguish. Additionally, the outcome of the data-driven models mostly depends on variations relating to the training dataset. It is difficult to include datasets with all possible variations in model training. This paper addresses both issues and proposes a novel hybrid technique where we extract novel physical features and data-driven features and then combine them to create an effective rain-streak removal strategy. The performance of the proposed algorithm has been tested in comparison to several relevant and contemporary methods using benchmark datasets. The experimental result shows that the proposed method outperforms the other methods in terms of subjective, objective, and object detection comparisons for both synthetic and real rain scenarios by removing rain streaks and retaining the moving objects more effectively. [ABSTRACT FROM AUTHOR]

Drowsy driving can significantly affect driving performance and overall road safety. Statistically, the main causes are decreased alertness and attention of the drivers. The combination of deep learning and computer-vision algorithm applications has been proven to be one of the most effective approaches for the detection of drowsiness. Robust and accurate drowsiness detection systems can be developed by leveraging deep learning to learn complex coordinate patterns using visual data. Deep learning algorithms have emerged as powerful techniques for drowsiness detection because of their ability to learn automatically from given inputs and feature extractions from raw data. Eye-blinking-based drowsiness detection was applied in this study, which utilized the analysis of eye-blink patterns. In this study, we used custom data for model training and experimental results were obtained for different candidates. The blinking of the eye and mouth region coordinates were obtained by applying landmarks. The rate of eye-blinking and changes in the shape of the mouth were analyzed using computer-vision techniques by measuring eye landmarks with real-time fluctuation representations. An experimental analysis was performed in real time and the results proved the existence of a correlation between yawning and closed eyes, classified as drowsy. The overall performance of the drowsiness detection model was 95.8% accuracy for drowsy-eye detection, 97% for open-eye detection, 0.84% for yawning detection, 0.98% for right-sided falling, and 100% for left-sided falling. Furthermore, the proposed method allowed a real-time eye rate analysis, where the threshold served as a separator of the eye into two classes, the "Open" and "Closed" states. [ABSTRACT FROM AUTHOR]

The rise in crime rates in many parts of the world, coupled with advancements in computer vision, has increased the need for automated crime detection services. To address this issue, we propose a new approach for detecting suspicious behavior as a means of preventing shoplifting. Existing methods are based on the use of convolutional neural networks that rely on extracting spatial features from pixel values. In contrast, our proposed method employs object detection based on YOLOv5 with Deep Sort to track people through a video, using the resulting bounding box coordinates as temporal features. The extracted temporal features are then modeled as a time-series classification problem. The proposed method was tested on the popular UCF Crime dataset, and benchmarked against the current state-of-the-art robust temporal feature magnitude (RTFM) method, which relies on the Inflated 3D ConvNet (I3D) preprocessing method. Our results demonstrate an impressive 8.45-fold increase in detection inference speed compared to the state-of-the-art RTFM, along with an F1 score of 92%,outperforming RTFM by 3%. Furthermore, our method achieved these results without requiring expensive data augmentation or image feature extraction. [ABSTRACT FROM AUTHOR]

Three-dimensional point cloud registration, which aims to find the transformation that best aligns two point clouds, is a widely studied problem in computer vision with a wide spectrum of applications, such as underground mining. Many learning-based approaches have been developed and have demonstrated their effectiveness for point cloud registration. Particularly, attention-based models have achieved outstanding performance due to the extra contextual information captured by attention mechanisms. To avoid the high computation cost brought by attention mechanisms, an encoder–decoder framework is often employed to hierarchically extract the features where the attention module is only applied in the middle. This leads to the compromised effectiveness of the attention module. To tackle this issue, we propose a novel model with the attention layers embedded in both the encoder and decoder stages. In our model, the self-attentional layers are applied in the encoder to consider the relationship between points inside each point cloud, while the decoder utilizes cross-attentional layers to enrich features with contextual information. Extensive experiments conducted on public datasets prove that our model is able to achieve quality results on a registration task. [ABSTRACT FROM AUTHOR]

The Vision Transformer (ViT) architecture has been remarkably successful in image restoration. For a while, Convolutional Neural Networks (CNN) predominated in most computer vision tasks. Now, both CNN and ViT are efficient approaches that demonstrate powerful capabilities to restore a better version of an image given in a low-quality format. In this study, the efficiency of ViT in image restoration is studied extensively. The ViT architectures are classified for every task of image restoration. Seven image restoration tasks are considered: Image Super-Resolution, Image Denoising, General Image Enhancement, JPEG Compression Artifact Reduction, Image Deblurring, Removing Adverse Weather Conditions, and Image Dehazing. The outcomes, the advantages, the limitations, and the possible areas for future research are detailed. Overall, it is noted that incorporating ViT in the new architectures for image restoration is becoming a rule. This is due to some advantages compared to CNN, such as better efficiency, especially when more data are fed to the network, robustness in feature extraction, and a better feature learning approach that sees better the variances and characteristics of the input. Nevertheless, some drawbacks exist, such as the need for more data to show the benefits of ViT over CNN, the increased computational cost due to the complexity of the self-attention block, a more challenging training process, and the lack of interpretability. These drawbacks represent the future research direction that should be targeted to increase the efficiency of ViT in the image restoration domain. [ABSTRACT FROM AUTHOR]

Saliency detection is a key research topic in the field of computer vision. Humans can be accurately and quickly mesmerized by an area of interest in complex and changing scenes through the visual perception area of the brain. Although existing saliency-detection methods can achieve competent performance, they have deficiencies such as unclear margins of salient objects and the interference of background information on the saliency map. In this study, to improve the defects during saliency detection, a multiscale cascaded attention network was designed based on ResNet34. Different from the typical U-shaped encoding–decoding architecture, we devised a contextual feature extraction module to enhance the advanced semantic feature extraction. Specifically, a multiscale cascade block (MCB) and a lightweight channel attention (CA) module were added between the encoding and decoding networks for optimization. To address the blur edge issue, which is neglected by many previous approaches, we adopted the edge thinning module to carry out a deeper edge-thinning process on the output layer image. The experimental results illustrate that this method can achieve competitive saliency-detection performance, and the accuracy and recall rate are improved compared with those of other representative methods. [ABSTRACT FROM AUTHOR]

Recently, computer vision-based methods have been successfully applied in many industrial fields. Nevertheless, automated detection of steel surface defects remains a challenge due to the complexity of surface defects. To solve this problem, many models have been proposed, but these models are not good enough to detect all defects. After analyzing the previous research, we believe that the single-task network cannot fully meet the actual detection needs owing to its own characteristics. To address this problem, an end-to-end multi-task network has been proposed. It consists of one encoder and two decoders. The encoder is used for feature extraction, and the two decoders are used for object detection and semantic segmentation, respectively. In an effort to deal with the challenge of changing defect scales, we propose the Depthwise Separable Atrous Spatial Pyramid Pooling module. This module can obtain dense multi-scale features at a very low computational cost. After that, Residually Connected Depthwise Separable Atrous Convolutional Blocks are used to extract spatial information under low computation for better segmentation prediction. Furthermore, we investigate the impact of training strategies on network performance. The performance of the network can be optimized by adopting the strategy of training the segmentation task first and using the deep supervision training method. At length, the advantages of object detection and semantic segmentation are tactfully combined. Our model achieves mIOU 79.37% and mAP@0.5 78.38% on the NEU dataset. Comparative experiments demonstrate that this method has apparent advantages over other models. Meanwhile, the speed of detection amount to 85.6 FPS on a single GPU, which is acceptable in the practical detection process. [ABSTRACT FROM AUTHOR]

Point cloud registration aims to find a rigid spatial transformation to align two given point clouds; it is widely deployed in many areas of computer vision, such as target detection, 3D localization, and so on. In order to achieve the desired results, registration error, robustness, and efficiency should be comprehensively considered. We propose a deep learning-based point cloud registration method, called DOPNet. DOPNet extracts global features of point clouds with a dynamic graph convolutional neural network (DGCNN) and cascading offset-attention modules, and the transformation is predicted by a multilayer perceptron (MLP). To enhance the information interaction between the two branches, the feature interaction module is inserted into the feature extraction pipeline to implement early data association. We compared DOPNet with the traditional method of using the iterative closest point (ICP) algorithm along with four learning-based registration methods on the Modelnet40 data set. In the experiments, the source and target point clouds were generated by sampling the original point cloud twice independently; we also conducted additional experiments with asymmetric objects. Further evaluation experiments were conducted with point cloud models from Stanford University. The results demonstrated that our DOPNet method outperforms these comparative methods in general, achieving more accurate and efficient point cloud registration. [ABSTRACT FROM AUTHOR]

Entire surface point clouds in complex objects cannot be captured in a single direction by using noncontact measurement methods, such as machine vision; therefore, different direction point clouds should be obtained and registered. However, high efficiency and precision are crucial for registration methods when dealing with huge number of point clouds. To solve this problem, an improved registration algorithm based on double threshold feature extraction and distance disparity matrix (DDM) is proposed in this study. Firstly, feature points are extracted with double thresholds using normal vectors and curvature to reduce the number of points. Secondly, a fast point feature histogram is established to describe the feature points and obtain the initial corresponding point pairs. Thirdly, obviously wrong corresponding point pairs are eliminated as much as possible by analysing the Euclidean invariant features of rigid body transformation combined with the DDM algorithm. Finally, the sample consensus initial alignment and the iterative closest point algorithms are used to complete the registration. Experimental results show that the proposed algorithm can quickly process large data point clouds and achieve efficient and precise matching of target objects. It can be used to improve the efficiency and precision of registration in distributed or mobile 3D measurement systems. [ABSTRACT FROM AUTHOR]

Nowadays, tool condition monitoring (TCM), which can prevent the waste of resources and improve efficiency in the process of machining parts, has developed many mature methods. However, TCM during the production of cutting tools is less studied and has different properties. The scale of the defects in the tool production process is tiny, generally between 10 μm and 100 μm for diamond tools. There are also very few samples with defects produced by the diamond tool grinding process, with only about 600 pictures. Among the many TCM methods, the direct inspection method using machine vision has the advantage of obtaining diamond tool information on-machine at a low cost and with high efficiency, and the method is accurate enough to meet the requirements of this task. Considering the specific, above problems, to analyze the images acquired by the vision system, a neural network model that is suitable for defect detection in diamond tool grinding is proposed, which is named DToolnet. DToolnet is developed by extracting and learning from the small-sample diamond tool features to intuitively and quickly detect defects in their production. The improvement of the feature extraction network, the optimization of the target recognition network, and the adjustment of the parameters during the network training process are performed in DToolnet. The imaging system and related mechanical structures for TCM are also constructed. A series of validation experiments is carried out and the experiment results show that DToolnet can achieve an 89.3 average precision (AP) for the detection of diamond tool defects, which significantly outperforms other classical network models. Lastly, the DToolnet parameters are optimized, improving the accuracy by 4.7%. This research work offers a very feasible and valuable way to achieve TCM in the manufacturing process. [ABSTRACT FROM AUTHOR]

Multi-object tracking is a significant field in computer vision since it provides essential information for video surveillance and analysis. Several different deep learning-based approaches have been developed to improve the performance of multi-object tracking by applying the most accurate and efficient combinations of object detection models and appearance embedding extraction models. However, two-stage methods show a low inference speed since the embedding extraction can only be performed at the end of the object detection. To alleviate this problem, single-shot methods, which simultaneously perform object detection and embedding extraction, have been developed and have drastically improved the inference speed. However, there is a trade-off between accuracy and efficiency. Therefore, this study proposes an enhanced single-shot multi-object tracking system that displays improved accuracy while maintaining a high inference speed. With a strong feature extraction and fusion, the object detection of our model achieves an AP score of 69.93% on the UA-DETRAC dataset and outperforms previous state-of-the-art methods, such as FairMOT and JDE. Based on the improved object detection performance, our multi-object tracking system achieves a MOTA score of 68.5% and a PR-MOTA score of 24.5% on the same dataset, also surpassing the previous state-of-the-art trackers. [ABSTRACT FROM AUTHOR]