Search

Magnetic resonance images are usually corrupted by noise during the acquisition process, which can affect the results of subsequent medical image analysis and diagnosis. This paper presents a denoising recurrent convolutional neural network for Brain MRI denoising. The proposed model consists of a one-level autoencoder architecture with a shortcut, in which the standard convolutional blocks are changed for a new recurrent convolutional denoising block. This block is based on the gated recurrent units combined with local residual learning, allowing us to filter the noisy image recursively. Additionally, we adopt global residual learning to directly estimate the corrupted image’s noise instead of the noise-free image. The proposed model requires less computation than other models based on neural networks and experimentally outperforms state-of-the-art models on clinical brain MRI datasets, particularly for high noise levels.

The lesion area segmentation technology in endoscopic images helps doctors discover and locate lesion areas, which is crucial for improving the survival rate of patients. In various tasks within the field of medical image segmentation, the fully convolutional network with U-shaped architecture has achieved great success, but the fully convolutional operation has limitations in obtaining global context information.The Segment Anything Model (SAM), the latest foundational model in computer vision, boasts powerful image feature extraction capabilities and can capture explicit long-range dependencies, but it cannot capture low-level detailed information. This paper proposes a dual-branch model, named FCSAM, which makes up for the respective shortcomings of the fully convolutional network and SAM in medical image segmentation. At the same time, this paper proposes a LayerNorm and low-rank-based(LNLoRA) fine-tuning strategy, which overcomes the problem that large models cannot be applied to small-scale medical image datasets and reduces the deployment and storage overhead of the model. In addition, we also design a boundary aware module and a boundary key point map generation algorithm to improve the boundary localization ability of this model, and significantly improve the segmentation performance of lesions. The experimental results show that the IOU, Dice coefficient and accuracy of the model are 79.3%, 88.1% and 93.2% respectively. These metrics show that the model proposed in this paper achieves superior performance to various competing methods in terms of segmentation accuracy and has great potential in clinical analysis and diagnosis.

Video Compression (VC) is a significant aspect of multimedia technology, in which the goal to minimize the size of video data, while also preserving its perceptual quality, for effective transmission and storage. Traditional approaches such as transform coding, predictive coding, and entropy coding are some of the much earlier discovered approaches in this area. VC is a challenging concept which plays a significant role in the effective transmission of data with low storage and minimum bandwidth requirements. However, the limited processing power, storage, memory, lower compression rate and lower resolution are some factors that impact the functionality and performance of VC. This survey aims to encompass a comprehensive review of present DL approaches for VC, especially the application of advanced DL-based Neural Network (NN) algorithms that are developed for solving the aforementioned challenges of VC. Adaptability of DL algorithms is exploited to enhance the potential quality of compressed videos and to positively influence lossless video compression outcomes. The DL approaches include Deep Neural Networks (DNN) methods such as Convolutional Neural Networks (CNN), Generative Adversarial Networks (GAN), Recurrent Neural Networks (RNN), Deep Recurrent Auto-Encoder (DRAE), etc. This survey examines the relationships, strengths as well as problem statements of DL-based compression approaches of VC. Furthermore, this survey also deliberates on datasets, hardware specifications, comparative analysis, and research directions. This survey embeds DL-based computer vision approaches, with hardware accelerators like GPU and FPGA, to minimize the complexity of in a model. This survey aims to overcome the limitations of VC, such as the varying effectiveness of specific encoder approaches, the challenges in utilizing hardware accelerators, low-resource devices, and difficulties in managing the large-scale databases. Integrating DL-based approaches with existing standard codecs remains a significant challenge. Ensuring compatibility, interoperability, and standardization is important for widespread adoption and integration. Enhancing the interpretability and control of DL approaches permit for better customization of compression settings, allowing the users to balance bit rate and quality according to their specific requirements. To gather relevant studies, widespread VC datasets are researched and utilized such as, Ultra-Video-Group dataset (UVG), Video Trace Library (VTL), etc. The selection criteria for this study of VC techniques and deep learning (DL) approaches are chosen to focus on the integration of DL with codecs, which is a primary research area of interest. This integration provides valuable insights into advanced DL applications in overcoming challenges associated with VC. Frameworks such as TensorFlow, Keras, PyTorch are utilized to classify the approaches according to their fundamental NN architectures.

Effective predicting of traffic flow is the key to fully utilizing the carrying capacity of roads and improving the travel experience. In order to overcome the randomness and volatility of traffic flow, a novel hybrid model for short-term traffic flow prediction based on spatio-temporal deep learning with considering associated factors selection is proposed. In this model, traffic flow with high spatial relevance is identified firstly with the use of the Pearson Product-Moment Correlation Coefficient (PPMCC), then associated factors affecting traffic flow are screened with the use of Random Forest (RF), finally, the results of the two steps mentioned above and the historical traffic flow data are used as input, and the Convolutional Neural Network (CNN) and Bidirectional Gated Recurrent Unit (BiGRU) are used to perceive the spatiotemporal characteristics of traffic flow and the hidden relationships between various factors and traffic flow to predict traffic flow. To test the performance of the proposed model, seven baseline models proposed in the existing literature are compared on publicly available datasets using four indexes to evaluate the performance of the models. In addition, we conducted an ablation study. The results showed that the proposed model has a 39% decrease in root mean square error and a 37% decrease in mean absolute error compared to the baseline models with the best performance.

Ensuring cutting tools are in optimal condition is essential for achieving peak machining performance, given their direct impact on both workpiece quality and process efficiency. However, accurately assessing wear on end mills, especially those with complex geometries, pose a significant challenge due to their reflective surfaces and varied wear patterns. Presented here is a novel method that addresses this challenge by employing a customized illumination unit in conjunction with a convolutional neural network (CNN) for end mill wear analysis. This innovative approach involves utilizing the specially designed illumination unit to capture high-quality images, enabling precise examination of material wear on helically shaped end mills. Notably, this method is tailored to illuminate reflective surfaces and represents a pioneering application in the realm of wear testing.We validate the viability of this approach by employing CNN-based models to segment wear on complex-shaped end mills coated with titanium carbonitride (TiCN) and titanium nitride (TiN). We achieved remarkable mean Intersection over Union (mIoU) results in wear detection on a test dataset: 0.99 for tool segmentation, 0.78 for abnormal wear, and 0.71 for normal wear segmentation.

Early detection of knee osteoarthritis is crucial because the damage in the knee joint is irreversible at the advanced stage. Medical images such as Magnetic Resonance Imaging plays an important role in knee osteoarthritis diagnosis as it provides excellent visualization of the osteoarthritis imaging biomarkers. Current clinical practice relies on manual inspection of the images which is very tedious, especially for 3D volumetric data. The overall aim of the study is to develop an efficient fully automated 3D segmentation model for segmenting multiple knee joint tissues from 3D Magnetic Resonance Imaging volumes. This study contributes by implementing hyperparameter optimization techniques to develop the optimal model for knee segmentation which will be beneficial for the detection of knee osteoarthritis. The model employs depthwise separable convolution for better computational efficiency. This paper presents an efficient model for knee bones and cartilages segmentation, modelled by Tree-of-Parzen-Estimators algorithm, which achieved an average dice score of 0.939 and a Jaccard index of 0.891. Our model outperformed 3D U-Net and 3D V-Net by approximately 7% and 6% respectively in terms of Dice Similarity Coefficient, with remarkably less computations, using the same dataset. The efficient model enhances the segmentation of the knee structures for better visualization, which contributes to a more accurate diagnosis in clinical practice. It also reduces the computational cost, allowing more possible adaptation of 3D neural networks in real-world clinical settings. Therefore, this work contributes to advance medical imaging and diagnostics while also holds the potential to improve clinical practice.

Owing to improvements in computing power, deep learning technology using convolutional neural networks (CNNs) has recently been used in various fields. However, using CNNs on edge devices is challenging because of the large computation required to achieve high performance. To solve this problem, pruning, which reduces redundant parameters and computations, has been widely studied. However, a conventional pruning method requires two learning processes, which are time-consuming and resource-intensive, and it is difficult to reflect the redundancy in the pruned network because it only performs pruning once on the unpruned network. Therefore, in this paper, we utilize a single learning process and propose an adaptive scaling method that dynamically adjusts the size of the network to reflect the changing redundancy in the pruned network. To verify the performance of each method, we compare the performance of the proposed methods by conducting experiments on various datasets and networks. In our experiments using the ImageNet dataset on ResNet-50, pruning FLOPs by 50.1% and 74.0% resulted in a decrease in top-1 accuracy by 0.92% and 3.38%, respectively, and improved inference time by 26.4% and 58.9%, respectively. In addition, pruning FLOPs by 27.37%, 36.84% and 46.41% using the COCO dataset on YOLOv7, reduced mAP(0.5-0.95) by 1.2%, 2.2% and 2.9%, respectively, and improved inference time by 12.9%, 16.9% and19.3%.

Dense multi-view image reconstruction has been a focal point of research for an extended period, with recent surges in interest. The utilization of multi-view images offers solutions to numerous challenges and amplifies the effectiveness of various applications including 3D reconstruction, de-occlusion, depth sensing, saliency detection, and identifying salient objects. This paper introduces an approach to reconstructing high-density light field (LF) images, addressing the inherent challenge of balancing angular and spatial resolution caused by limited sensor resolution. We introduce an innovative approach to reconstructing LF images through a CNN-based network that combines spatial and epipolar features in both initial and deep feature extraction phases. Our network utilizes angular information during upsampling and employs dual feature extraction to effectively analyze horizontal and vertical epipolar data. Weight sharing within the CNN block between horizontal and vertically transposed stacks enhances quality while preserving model compactness. The outcomes of experiments carried out on real-world and synthetic datasets demonstrate the effectiveness of our method, showcasing its superior performance in both inference speed and reconstruction quality when compared to state-of-the-art (SOTA) techniques.

Many approaches to automated epileptic seizure detection share a common challenge — the trade-off between recall and precision. This study aims to develop a novel approach for reducing false positive predictions in seizure detection tasks applied to real-world EEG recordings. We propose a multi-stage modeling framework, for which the novelty lies in combination of traditional machine learning outlier detection with state-of-the-art convolutional neural networks. Our dataset includes raw epileptic EEG data directly from the hospital. Continuous wavelet analysis is employed for EEG preprocessing and feature extraction. We evaluated the performance of the proposed two-stage algorithm, and it demonstrated a slight decrease in recall but a significant improvement in precision in comparison to machine-learning-only or neural-network-only algorithms. We hypothesize that this finding aligns well with our previous research and relates to the fundamental properties of epileptic EEG, including the extreme behavior of seizures. Finally, we propose a potential practical application of the developed approach within a clinical decision support system.

The Joint Video Experts Team (JVET) has explored neural network-based video coding (NNVC) and is trying to introduce NNVC into the versatile video coding (VVC). In NNVC, the NN-based in-loop filter is the most active area, which is very close to deployment of software. Recent NN-based in-loop filters start adopting Transformer to capture context information, but it causes a remarkable increase of complexity to about 1000 kMAC/Pixel. In this paper, we propose a low complexity NN-based in-loop filter for VVC based on convolution and Transformer, named ConvTransNet. ConvTransNet adopts a pyramid structure in feature extraction to capture both global contextual information and local details at multiple scales. Moreover, ConvTransNet combines convolutional neural network (CNN) and Transformer into the in-loop filter. CNN captures local features and reduces compression artifacts in an image, while Transformer captures long-range spatial dependency and enhances global structures in an image. Thus, ConvTransNet enables the NN-based in-loop filter to reduce compression artifacts and enhance visual quality in an image. In ConvTransNet, we use grouped convolutions in CNN and depthwise convolutions in Transformer to reduce the network complexity. Therefore, ConvTransNet successfully captures both local spatial structure and global contextual information in an image and achieves outstanding performance in terms of BD-rate and complexity. Experimental results show that the proposed NN-based in-loop filter based on ConvTransNet achieves average {6.58%, 23.02%, 23.04%} and {8.18%, 22.67%, 22.00%} BD-rate reductions for {Y, U, V} channels over VTM_11.0-NNVC_2.0 anchor under AI and RA configurations, respectively.

With the continuous development of wireless technology, automatic modulation recognition plays an increasingly prominent role in military and civilian fields. However, the complex communication environment and diversified strategy bring challenges to modulation signal recognition. Therefore, an automatic modulation signal recognition technique combining truncated migration processing and convolutional neural network is proposed. Then multi-task learning is used to optimize and distinguish easily confused modulated signals. Three datasets, RadioML2016.10A, RadioML2016.10B and RadioML2016.04C are used for experiments. The results show that the modulation signal automatic recognition method proposed in this paper has good recognition accuracy. When the signal-to-noise ratio was 14dB, it reached a maximum of 95.46%. In addition, the Floating Point Operations of the proposed method were 1.71G, the number of parameters was 25636712, and the operation time was 228s, which showed that the method was light in weight and high in calculation efficiency. The optimized recognition technology can effectively distinguish three groups of easily confused signals, and the recognition rate is as high as 100%, and the minimum is not less than 90%. The proposed modulated signal automatic recognition technology has achieved remarkable results in improving the recognition accuracy and processing efficiency, which provides a strong technical support for the development of wireless communication technology, and also provides a new tool for researchers and engineers in related fields.

Abnormal gait patterns are a common feature of Cerebral Palsy, a neurodevelopmental disease for which early identification is essential for treatment. In the proposed research, a novel methodology is provided for classifying abnormal gait patterns in children with Cerebral Palsy, using gait analysis as a diagnostic tool. To improve gait classification accuracy and efficiency, a hybrid model of Convolutional Long Short-Term Memory (ConvLSTM) model and Generative Adversarial Network (GAN) is used in the suggested technique. The proposed study concentrated on temporal signal data, using hypothetical planes with minimal regard for anatomical indicators. The reduction technique enables a more efficient and successful gait analysis. Heatmap images were created from the selected temporal data. GAN generated images were added to the dataset in order to overcome the problems caused by class imbalance and guarantee a thorough depiction of abnormal gait patterns. In the proposed work, a ConvLSTM-based model with a batch size of 32, training as well as validation datasets were evaluated over a period of 50 epochs. The effectiveness of the suggested model was compared to other models such as Gated Recurrent Unit, Convolutional Neural Network, and Long Short-Term Memory model that were trained using the same input data. Our suggested ConvLSTM model produced an impressive accuracy of 91.8% and a loss of 0.42. The Convolutional Long Short Term Memory model performed better than the other models when compared based on a number of criteria, including accuracy, precision, recall, and F1-score. The performance measures demonstrate how well our method works to classify the abnormal gait in kids with Cerebral Palsy.

Human monitoring using radar systems operating in the GHz regime has generated significant interest as a result of the increasing availability of commercial radar systems. These sensors offer all weather performance, the ability to measure range and velocity, and the protection of anonymity. However, visually inferring activities present in radar data is often challenging without prior knowledge. Here, we address this by implementing a radar-to-pose system that converts the raw radar data into human poses, such that human forms can be identified and activities monitored. In comparison to prior works, we place our radar in an elevated position, more in line with the placement of existing real world monitoring systems e.g. cameras, or emerging systems, e.g. quadcopters. We present an ensemble predictor network and apply it to a number of human poses of increasing complexity, reporting accuracies in excess of 90%, and verify the generalizable nature of our approach with unseen validation data. We perform an in depth explainability analysis, exploiting the unique mappings created by our radar placement and network structure to confirm that the network is making rational predictions based on the true location of limbs.

Temperature forecasting of grain in storage is crucial for timely granary temperature control, mitigating adverse effects of extreme temperatures on grain quality. Traditional machine learning methods struggle with stability and high error rates in grain storage temperature forecasting, while deep learning models are more accurate but time-consuming and have heavy parameters. To address these problems, an improved model with light weight and good accuracy is proposed in this paper, which broad learning network is combined with one-dimensional convolution module and multi-head self-attention mechanism (BLN-1DCNN-MHSA). Firstly, we employ a one-dimensional convolution module at the feature nodes of the model to extract local temporal correlations, compensating for temporal sequence learning limitations of the BLN. Secondly, a multi-head self-attention mechanism at the enhancement nodes to captures important features dependencies and global temporal correlations. Lastly, our model achieves better prediction through enhanced representation ability of model nodes. The results with real grain storage temperature data demonstrate that the RMSE, MAPE, and MAE of the proposed model are 0.341, 0.54%, 0.28, respectively, which represent more than 2 times improvement in accuracy compared to the BLN, and it also reduces training time by more than 90% compared with LSTM and Transformer models. Additionally, the generalization and robustness of the improved approach are demonstrated through promising results in a classification experiment on the MNIST dataset. In general, the model provides a certain feasibility for early warning of grain storage risks by predicting its temperature trends.

Deep learning methods based on convolutional neural networks can identify subtle disease features in plant leaves, thereby improving the accuracy and efficiency of plant leaf disease detection. Traditional convolutional neural network models have more parameters, lower training efficiency, and require a large amount of computing resources. A TTN-MobileNetV2 neural network model based on memristors for plant leaf disease detection is proposed in this paper. Firstly, integrate the Triplet attention module into the backbone structure of the network to capture local features, and utilize Cross-Norm and Self-Norm(CNSN) normalization techniques to enhance the generalization robustness under distribution changes. In addition, a Mish activation function with enhanced nonlinear characteristics was introduced to improve the accuracy of neural network detection. Experimental results on the Plant Village and Rice leaf disease datasets showed identification accuracies of 99.03% and 99.16%, respectively. On this basis, using the MemTorch simulation environment, the weights of all convolutional layers and fully connected layers in the convolutional neural network are mapped to the conductance values of the memristors in the cross array of memristors, completing the implementation of the memristor TTN-MobileNetV2 network. The performance of the memristor network was tested using two types of memristor models: linear ion drift model and data-driven Verilog-A RRAM. The recognition accuracy losses of the TTN-MobileNetV2 memristor network corresponding to the two memristor models were 0.32, 0.34, and 0.52, 0.61, respectively. So the memristor convolutional neural network can meet the performance requirements of plant leaf disease recognition and has inherent advantages of high speed and low power consumption.

Partial discharge (PD) faults often occur as a result of breakdowns in the insulation layer of insulated overhead conductors. PD faults can cause serious problems such as power outages or electrical fire accidents. In this paper, a new PD detection system based on spectral analysis, spectrogram analysis, deep learning algorithms, minimum redundancy-maximum relevance (mRMR) and ensemble machine learning (EML) is presented. In the process of extracting distinctive features in the frequency dimension, 1D spectral data and 2D spectrum data based on frequency-time are obtained from a raw PD signal. Fourier transform-based three-power spectral density analyzes and one spectrogram analysis are performed. Deep features are obtained by using pre-trained 1D convolutional neural network models for 1D spectral data and the pre-trained ResNet-50 model for 2D spectrogram data. The most effective features are determined by applying mRMR feature selection analysis to the obtained deep features. In the last stage, PD detection is performed by applying the selected deep features to the EML classifier. The performance of the proposed PD detection system are evaluated with the VSB common data set. According to the experimental results, the proposed deep feature approach based PD detection system has very high performance.

Histopathological image classification stands as a cornerstone in the pathological diagnosis workflow, yet it remains challenging due to the inherent complexity of histopathological images. Recently, transformers and convolutional neural network (CNN) - based deep models have shown promising results in the automatic histopathology image classification. Transformers excel at capturing global dependencies within the image content, while CNNs effectively extract local features. In this study, we introduce the CViTS-Net model, a novel deep learning architecture that combines CNNs with vision transformer (ViT), enhanced by innovative skip connections. This fusion allows our model to capture both local and global dependencies within histopathological images. Extensive experiments were conducted, including holdout validation and cross-validation, comparing CViTS-Net with several state-of-the-art CNN, ViT, and attention-based methods on the Chaoyang histopathology dataset. Furthermore, we evaluated the model’s generalizability and robustness by testing it on large and diverse datasets such as lymphoma dataset and invasive ductal carcinoma (IDC) dataset. Our model achieves remarkable classification accuracy of 96.06% on the Chaoyang dataset, 99.61% accuracy on the lymphoma dataset, and 95% accuracy on IDC dataset, surpassing state-of-the-art deep learning models while maintaining superior efficiency. The CViTS-Net model showcases outstanding classification performance, underscoring its potential to significantly aid pathologists in histopathological diagnosis.

Polyp segmentation is vital for early detection and treatment of colorectal cancer, significantly improving patient prognosis. This paper proposes an efficient and precise polyp segmentation model called the Partial Decoder Localization and Foreground-Background Refinement Network (PDLFBR-Net), which simulates the human object recognition process. Specifically, PDLFBR-Net comprises three key modules: the Cross-level Attention-enhanced Fusion Module (CAFM), the Position Recognition Module (PRM), and the Foreground-Background Refinement Module (FBRM). The CAFM enhances feature representation by fusing information from adjacent levels, providing more discriminative features. The PRM module simulates the human recognition process by using a partial decoder to locate potential polyp tissues from a global perspective. Subsequently, the FBRM is used to perform specific recognition, gradually refining the initial prediction results through foreground and background focusing to achieve precise recognition. Extensive experiments demonstrate that the proposed PDLFBR-Net model significantly outperforms existing state-of-the-art models on five challenging datasets. On the Kvasir-SEG benchmark dataset, the mean Dice and mean IoU values reached 93.7% and 89.5%, respectively, which represents an improvement of 0.4% and 0.6% compared to the best-performing state-of-the-art (SOTA) method.

Clinical methods for dementia detection are expensive and prone to human errors. Despite various computer-aided methods using electroencephalography (EEG) signals and artificial intelligence, a reliable detection of Alzheimer’s disease (AD) remains a challenge. The existing EEG-based machine learning models have limited performance or high computation complexity. Hence, there is a need for an optimal deep learning model for the detection of AD. This paper proposes a low-complexity EEG-based AD detection CNN called LEADNet to generate disease-specific features. LEADNet employs spatiotemporal EEG signals as input, two convolution layers for feature generation, a max-pooling layer for asymmetric spatiotemporal redundancy reduction, two fully-connected layers for nonlinear feature transformation and selection, and a softmax layer for disease probability prediction. Different quantitative measures are calculated using an open-source AD dataset to compare LEADNet and four pre-trained CNN models. The results show that the lightweight architecture of LEADNet has at least a 150-fold reduction in network parameters and the highest testing accuracy of 99.24% compared to pre-trained models. The investigation of individual layers of LEADNet showed successive improvements in feature transformation and selection for detecting AD subjects. A comparison with the state-of-the-art AD detection models showed that the highest accuracy, sensitivity, and specificity were achieved by the LEADNet model.

Skin lesion diagnosis in medical image analysis is still a difficult task. A frequent kind of cancer known as skin cancer affects people worldwide and can be fatal. As a result, early and accurate diagnosis is crucial for finding skin cancer patients. One of the most recent technologies for detecting skin cancer is dermatoscopy. For proper treatment of skin cancer, an early diagnosis is required. The early stages of skin lesions are identical, making manual diagnosis difficult. Therefore, creating computer-aided methods for classifying skin lesions will assist dermatologists in detecting skin lesions earlier and treating them more successfully. Current research indicates a significant potential for the classification of skin lesions using deep learning networks. However, issues including unbalanced datasets, poor contrast lesions, and the extraction of pointless or duplicate features still need to be resolved. This study aims to propose a transfer learning-based ensemble model for more accurate classification results. InceptionV3 and Xception are employed to build an ensemble model for classifying skin lesion images known as IncepX-Ensemble. A traditional data augmentation method was employed for the HAM10000 dataset to address the imbalanced data. This technique mitigates the class imbalance by incorporating data augmentation, enhancing model accuracy. At the experiment’s outset, we utilized the original imbalanced data, and subsequently, balanced data were employed for the proposed model. Training and testing accuracy was achieved at 86% on the imbalanced dataset. A balanced data augmentation dataset yielded 98% training and 98% test accuracy rates. We evaluated the output of the proposed model against outputs from various transfer learning models using both the original and balanced datasets.

Accurate models of zenith tropospheric delay (ZTD) is crucial in meteorology as well as in navigation and positioning. In this study, we employ Convolutional Neural Network (CNN) and Random Forest (RF) models to establish six direct or compensation models for estimating ZTD in Chinese mainland and surrounding areas. The modeling process utilizes ZTD data from 205 stations spanning the period 2013 to 2018. Model validity is assessed using ZTD data from 202 stations in 2019. Comparative analysis, considering the overall Root Mean Square Error (RMSE), is conducted between these newly proposed CNN/RF-based models and Saastamoinen, A&N, GPT3 and RF-based models constructed by the methods presented in the previous study (ZTD-RF1, ZTD-RF3). The results demonstrate the superiority of the six CNN/RF-based models over the previously proposed models. In general, compensation models exhibit an improvement over direct models, and models incorporating meteorological parameterisation outperform models without such parameterisation. When the meteorological data are available, our proposed model provided a good representation of the instability of water vapour pressure in the ZTD, especially in monsoon climates. The optimal model is identified as the RF-based compensation model (ZTD-RF4). The ZTD-RF4 model achieves an overall RMSE of 3.24 cm, representing a 29.47% reduction of the RMSE compared to the Saastamoinen model (4.60 cm), a 26.75% reduction compared to the A&N model (4.43 cm), and slightly superior to the ZTD-RF3 model (3.28 cm). When the meteorological data are unavailable, the optimal choice is the CNN-based compensation model (ZTD-CNN2), which exhibits an overall RMSE of 4.21 cm, indicating a 7.89% reduction compared to the GPT3 model (4.57 cm) and significantly superior to the ZTD-RF1 model (4.34 cm). In contrast to current machine learning (ML)-based ZTD calculation models, we introduce the idea of compensation based on traditional models and a new CNN structure is constructed, which all proved to be capable of better performance in ZTD modeling.

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.

With the widespread availability and utilization of 3D meshes across various applications, the need for accurately assessing their visual quality has become increasingly important. Despite the significance of this task, the literature offers few No-Reference (NR) approaches for evaluating the visual quality of 3D meshes. In response to this gap, this paper proposes a novel NR approach tailored specifically to score the quality of 3D meshes. After rendering a 3D mesh into 2D views and patches, a pre-trained convolutional neural network automatically extracts deep features from. These features are then employed in a Multi-Layer Perceptron regressor to predict the quality score of the rendered images. The obtained scores are combined with their corresponding BRISQUE scores, and an additional MLP regressor is used to predict the final score. We present experimental results demonstrating the effectiveness and robustness of our approach across a diverse range of 3D mesh datasets. Comparative analyses with existing NR methods underscore the superior performance and versatility of the proposed approach. Overall, this paper contributes to the advancement of NR techniques for assessing 3D mesh quality, offering a valuable tool for researchers, practitioners, and developers working with 3D models across various domains.

Suspicious Human activity recognition (SHAR) is crucial for improving surveillance and security systems by recognizing and reducing possible hazards in different situations. This study focuses on the task of precisely identifying potentially suspicious human behaviour by utilizing an innovative approach that harnesses advanced deep learning methods. Despite the abundance of research on the subject of SHAR, current methods frequently need to be revised with restricted levels of precision and efficiency. This research aims to address these constraints by presenting a thorough methodology for detecting and recognizing suspicious human activities. By rigorously collecting and preparing data, as well as training models, we aim to solve the issue of inaccurate and inefficient activity recognition in surveillance systems. By utilizing Convolutional Neural Networks (CNNs) and deep learning structures, such as the proposed time-distributed CNN model and Conv3D model, we attain notably enhanced accuracy rates of 90.14% and 88.23%, respectively, surpassing current research approaches. Moreover, the efficacy of our approach is illustrated by conducting prediction experiments on previously unreported test data and YouTube videos. Through the process of evaluating the trained models on unseen test data, we ascertain their accuracy and ability to apply learned knowledge to new situations. Moreover, the algorithms are utilized to predict dubious human conduct in a YouTube video, demonstrating their practical usefulness in real-life surveillance situations. The results of this study have important consequences for improving surveillance and security systems, allowing for better identification and reduction of possible dangers in various settings. Our methodology enhances the precision and effectiveness of SHAR, advancing the construction of more resilient and dependable surveillance systems and ultimately strengthening public safety and security.

As the economy and society continue to develop, the types of underwater targets are becoming increasingly diverse, accompanied by a corresponding increase in environmental noise. This noise presents a significant challenge for the recognition of underwater target radiation noise, as it can result in the loss of effective information during the noise processing stage. This, in turn, reduces the accuracy of underwater target radiation noise recognition. This paper proposes an underwater target radiation noise recognition method based on the enhanced image method. The method transforms the original underwater target radiation noise signal into an enhanced image, builds a convolutional neural network with the enhanced image as an input, and uses the convolutional neural network’s ability to classify images to recognise and classify underwater target radiation noise. The experimental results demonstrate that the training time of the method described in this paper is longer than that of the traditional machine learning method, yet the recognition and classification accuracy is significantly higher.

Pavement distress detection is crucial in road health assessment and monitoring. However, there are still some challenges in extracting pavement distress based on deep learning: such as insufficient segmentation, extraction errors and discontinuities. In this paper, we propose DARNet, a network for pavement distress extraction. A Distress Aware Attention Module (DAAM) is proposed to solve the problem of discontinuity in distress extraction due to inaccurate recovery of distress pixels during upsampling. Based on the characteristics of distress morphology, a Refinement Extraction Module (REM) is designed to effectively capture horizontal and vertical pavement damage features by fusing high-level and low-level features, which improves the accuracy of the model in extracting details of pavement damage information. Finally, a Weighted Cross-Entropy Loss function (WCEL) is introduced to assign weights according to the distance of the pixel point to the boundary of the distress, which solves the problem that the traditional cross entropy function treats each pixel point equally. We also propose a set of pavement distress datasets LNTU_RDD_GS, and the experimental results show that DARNet can reach 82.68% mIoU and 90.13% F score in the datasets in this paper, 80.63% mIoU and 88.35% F score in the four public datasets.

With the rapid advancement of computer vision technology, various deepfake tools for generating deceptive images have emerged. Generative Adversarial Networks (GANs) can create various deceptive media streams, including images, audio, and video, leading to numerous societal challenges. Palmprint recognition technology has recently been applied in financial identity verification, particularly in confirming transactions across various banking platforms. Manipulating critical financial transactions or generating malicious images to deceive authentication processes can result in significant disruptions. Convolutional Neural Networks (CNNs) are considered practical tools. We propose the implementation of a Dual Cascade Convolutional Neural Network (DC-CNN) algorithm that utilizes a dual-channel technique. This approach involves two networks that train one subnetwork and then apply the same configuration to the other. The feature vectors are combined, the fake inputs can be identified. This dual-channel technique is particularly effective for detecting forged images. Our approach involves comparing various CNN architectures, such as MesoNet, MesoInceptionNet, and Dense CNN (D-CNN), within the framework of GAN methods, such as Wasserstein GAN (WGAN) and Cycle GAN. In our experiments, DC-CNN demonstrates favorable results in detecting fake palmprints based on WGAN and cycle GAN. Specifically, for WGAN-based fake palmprints, the model achieved the weighted precision of 90.83%, weighted recall of 90.20%, weighted F1 scores of 89.92% and accuracy of 90.20%. In the case of Cycle GAN-based fake palmprints, the model exhibited the weighted precision of 87.86%, weighted recall of 87.91%, weighted F1 scores of 87.85% and accuracy of 87.91%. Therefore, DC-CNN emerges as a promising approach in the fields of deepfake palmprint detection and identity verification.

In this paper, a table lookup-based computing technique is proposed to perform convolutional neural network (CNN) inference without multiplication, and its FPGA implementation is demonstrated as a proof-of-concept. Conventionally, the hardware specific to the lookup-based dot product approximation (LDA) has been proposed to achieve energy-efficient CNN computations for edge AI applications. However, it has not been applied to complicated AI tasks, such as ImageNet image classification, because LDA degrades the inference accuracy, especially in complicated tasks. Therefore, a new LDA technique using residual vector quantization, called RLDA, is proposed in this study. By adopting the proposed RLDA to CNN, we can achieve an inference accuracy degradation of below 5% for ResNet-18 model on ImageNet 1000 classification. In addition, the proposed RLDA-based CNN accelerator is implemented on a ZCU104 evaluation board, which includes a Zynq UltraScale+ FPGA and DDR4 DRAM. We compare the processing time of the proposed accelerator on ZCU104 with that of NVIDIA Jetson AGX Orin and reveal that the processing time of the proposed accelerator is comparable to that of Jetson Orin in the lower layers of the ResNet-18 model. Finally, the computational performance of the proposed accelerator is compared with conventional FPGA-based accelerators. The proposed accelerator does not require a digital signal processor (DSP), which is often available in modern FPGAs. We demonstrate that the proposed accelerator achieves a performance more than four times higher than that of the conventional FPGA-based accelerator without DSP.

Cervical spine diseases, encompassing conditions like spondylolisthesis, disc degeneration, and cervical spinal stenosis, stand as significant contributors to global disability. Precise classification of these conditions is paramount for effective medical diagnosis. This paper introduces an innovative methodology aimed at addressing the limitations of traditional convolutional neural networks (CNNs) and pretrained models in this domain. We propose a novel approach dubbed Inv-AlxVGGNets, which leverages concatenated pretrained architectures AlexNet and VGG, augmented by involutional neural networks and residual layers. Unlike conventional CNNs that are location-specific and channel-agnostic, involutional neural networks offer enhanced Adaptability to diverse visual patterns in medical images. Focusing on a four-class cervical spine disease classification task utilizing MRI images, our study evaluates the performance of Inv-AlxVGGNets (AlexNet with INN and VGG with residual layer models) as well as machine learning algorithms. Our findings demonstrate superior performance in terms of accuracy, precision, recall, and AUC ROC values. Notably, Inv-AlxVGGNets achieves an impressive 98.73% accuracy on the testing set and 99.78% on the training set, underscoring its potential for precise cervical spinal disease classification. In a comparative analysis, we highlight that conventional CNNs entail over 133 million parameters, whereas Inv-AlxVGGNets require less than 8 million parameters, rendering them more efficient and resource-friendly. This reduced parameter count is particularly advantageous in resource-constrained scenarios, where computational resources and datasets may be limited. The promising results position Inv-AlxVGGNets as a valuable tool for precise cervical spine disease classification, offering implications for enhancing patient care in resource-constrained settings.

Image and video classification are distinct tasks in computer vision. Three-dimensional convolutional neural networks (3D CNNs) are commonly employed for video classification, while two-dimensional convolutional neural networks (2D CNNs) are more suitable for image classification. To enable image and video recognition to adopt the same network-3D CNNs, we propose a transform domain learning approach for image recognition utilizing the video recognition model 3D CNNs. The transform domain learning not only permits the use of RGB images as input for 3D CNNs but also allows for the input of orthogonally transformed image sequences into the networks. Furthermore, randomly transformed images can be fed into the networks, where the random transformation is a customized arbitrary transformation. The standard 3D CNNs can be seamlessly applied to both images and videos. The experiments show that whether orthogonal transformation or random transformation is used as the input, 3D CNNs can effectively classify images. Compared to 2D CNNs, the data after orthogonal transformation does not reduce the accuracy. To unify the tasks of image and video classification, the image dataset Caltech101 and the video dataset UCF101 are mixed, and 3D CNNs are used to recognize images in the transform domain. The results illustrate that mixed and individual training produce almost the same recognition effect. Additionally, it can be observed that directly transferring the video pre-training model to the image classification task can significantly enhance the performance.

With the popularity of edge computing devices and increasing complexity of convolutional neural network (CNN) models, the need for model compression and acceleration has become increasingly urgent. As an effective model compression technique, CANDECOMP/PARAFAC (CP) decomposition relies heavily on the preset rank for its compression effectiveness. However, no direct algorithm is currently available for determining the optimal tensor rank. Therefore, a novel method for CP rank selection based on deep reinforcement learning is proposed. This method utilizes the DecG single-player game framework based on the Deep Deterministic Policy Gradient (DDPG) algorithm to achieve automation and intelligence in rank selection. In this process, a pre-trained model is introduced, which fuses and reshapes several historical tensors as network inputs. Additionally, a hybrid greedy strategy based on singular value decomposition (SVD) was designed in the exploration phase to enhance the efficiency of finding ideal rank selection results. This method can automatically determine the rank according to the weight tensor characteristics of the convolution layer and optimize the compression efficiency and performance of the model. In addition, a compression efficiency index is developed to visually demonstrate the performance of the various compression methods. Finally, on the CIFAR-10 and CIFAR-100 datasets, CP decomposition experiments are conducted on various convolutional neural network models, and the decomposed models undergo iterative fine-tuning for retraining. The experimental results show that the rank values determined by the DecG method achieve significant optimization and enhancement in the compression efficiency of the models compared to other methods, exhibiting strong robustness.

This paper presents a Spatio-Temporal Road Inference for a KEeping NETwork (StrikeNet), aimed at enhancing Road Lane Reconstruction (RLR) and lateral motion control in Autonomous Vehicles (AV) using deep neural networks. Accurate road lane model coefficients are essential for an effective Lane Keeping System (LKS), but the traditional vision system often fails in situations where lane markers are absent or faint and cannot be properly recognized. To overcome this, a driving dataset was restructured, combining road information from a vision system and forward images for spatial training of RLR. Sequential spatial learning outputs were then processed with in-vehicle sensor data for temporal inference via Long Short-Term Memory (LSTM). The StrikeNet was rigorously tested in both typical and uncertain driving environments. Comprehensive statistical and visualization analyses were conducted to evaluate the performance of various RLR methods and lateral motion control strategies. Remarkably, the RLR demonstrated its capability to derive reliable road coefficients even in the absence of lane markers. Upon performance comparison with four alternative techniques, our method yields the lowest error and variance between human steering inputs and the control input. Specifically, under high and low lane quality conditions, the proposed method maximally reduced the control input error by up to 72% and 66%, respectively, and decreased the variance by 54% and 94%, respectively. The findings highlight StrikeNet’s effectiveness in bolstering the fail-operational performance, and reliability of lane-keeping or lane departure warning systems in autonomous driving, thereby enhancing control continuity and mitigating path deviation-induced traffic accidents.

Morphological differentiation between myeloblasts and monoblasts is pivotal for the majority of acute myeloid leukemia (AML) diagnosis in clinical settings. Manual morphology-based classification of blasts encounters challenges due to the limited differentiation of these bone marrow cells (BMC) of early stages. Hence, the utilization of artificial intelligence is essential to assist in the classification of these cells. 4001 single-cell images of monoblasts and myeloblasts were collected from Taizhou Hospital to form the BMC-1 dataset. The main novelties and features of the proposed method are as follows: 1) A maximum connected domain extraction method grounded in the watershed algorithm is introduced to efficiently eliminate stained impurity cells from single cell images. 2) A dynamic focal loss is introduced to gradually focus on difficult-to-classify samples as the training progresses. 3) A dual-channel convolutional block attention network (DCCBANet) is introduced to enhance feature extraction. It employs attention mechanisms to focus on key features, utilizing ordinary convolution for local feature extraction, dilated convolution for global feature extraction, and a decreasing dilation rate sequence to preserve detailed information. A macro F1-score (macro_F) of 96.8% is achieved on the BMC-1 validation dataset. Additionally, the presence of multiple differentiation types of granulocytes pose difficulties in granulocytes distinction. To further validate the efficacy of our proposed method on multi-classification tasks, we collected 6626 granulocyte single-cell images from Taizhou Hospital. Augmenting the dataset with 2441 granulocyte single-cell images from the public dataset BM_cytomorphology addressed sample shortages, forming the BMC-2 dataset. We applied our model to the BMC-2 dataset for experiments and ultimately achieved a macro_F of 87.49%. Our proposed method effectively distinguishes monoblasts and myeloblasts and excels in the classification of granulocytes.

Texture perception plays a vital role in various fields, from computer vision to geology, influencing object recognition, image segmentation, and rock classification. Despite advances in convolutional neural networks (CNNs), their effectiveness in texture-based classification tasks, particularly in rock classification, still needs exploration. This paper addresses this gap by evaluating different CNN architectures using diverse publicly available texture datasets and custom datasets tailored for rock classification. We investigated the performance of 38 distinct models pre-trained on the ImageNet dataset, employing both transfer learning and fine-tuning techniques. The study highlights the efficacy of transfer learning in texture classification tasks and offers valuable perspectives on the performance of different networks on different datasets. We observe that while CNNs trained on datasets like ImageNet prioritize texture-based features, they face challenges in nuanced texture-to-texture classification tasks. Our findings underscore the need for further research to enhance CNNs’ capabilities in texture analysis, particularly in the context of rock classification. Through this exploration, we contribute insights into the suitability of CNN architectures for rock texture classification, fostering advancements in both computer vision and geology.

Soil erosion poses significant ecological and economic challenges, necessitating precise and effective identification methods. Traditional models frequently overlook the intricate relationships between erosion factors and multispectral remote sensing data. To enhance these traditional methods, a novel dual-input gated fusion Convolutional Neural Network (CNN) has been developed, integrating channel and spatial attention mechanisms. This innovative model strengthens the connection between multispectral images and erosion factors, improving the accuracy and generalizability of erosion predictions. The model utilizes data collected from unmanned aerial vehicles (UAVs) equipped with high-precision multispectral sensors. By processing both spectral images and erosion factor data, the model effectively captures complex soil spatial distributions. The dual-input gated fusion mechanism allows the network to extract high-level semantics while suppressing redundant information, ensuring robust performance even in heterogeneous terrains. Experimental results indicate that this framework significantly enhances the performance of traditional models, providing superior predictions for small and medium-sized areas. Experimental results indicate that the presented framework can achieve better accuracy (96.92%) compared with other machine learning approaches, such as Random Forest (89.64%), VGGNET (91.52%), and RESNET (90.18%). Moreover, the proposed method can improve accuracy by 26.59% compared to the traditional RUSLE model. This improvement is critical for applications in ecological restoration and sustainable development. The integration of deep learning techniques with UAV-based data collection offers a powerful tool for environmental monitoring and management.

In semantic segmentation tasks for remote sensing images, effective feature extraction acts as the most important foundation. As a result, this paper proposes a novel semantic segmentation method for remote sensing images through adaptive scale-based convolution neural network. Firstly, it uses an encoder to extract features from remote sensing images, and utilizes attention mechanisms to control the information flow in the neural network. This is expected to reduce the impact between different scales. Then, a semantic segmentation for updating scale weights is established, and a scale-adaptive convolutional neural network is constructed. The upsampling unit is improved to increase the resolution to original image level, in order to better identify smaller targets. For large-sized problems, pyramid-like processing is used to segment images from multiple scales, and results are finally merged. Besides, we also make some experiments on ISPRS Potsdam dataset, UC Merced dataset, and DeepGlobe dataset, in order to make performance evaluation. The research shows that the maximum pixel accuracy of the proposed remote sensing semantic segmentation method is increased to 86.18%, the average value of the semantic segmentation task is up to 63.72, and the fastest running speed is up to 9.16FPS. In other words, the method proposed in this study has more accurate processing results and better stability.

Malware is a malicious software intended to cause damage to computer systems. In recent times, significant proliferation of malware utilized for illegal and malicious goals has been recorded. Several machine and deep learning methods are widely used for the detection and classification of malwares. Image-based malware detection includes the usage of machine learning and computer vision models for analyzing the visual representation of malware, including binary images or screenshots, for the purpose of detecting malicious behaviors. This techniques provides the potential to identify previously hidden or polymorphic malware variants based on the visual features, which provide a further layer of defense against emerging cyber-attacks. This study introduces a new Snake Optimization Algorithm with Deep Convolutional Neural Network for Image-Based Malware Classification technique. The primary intention of the proposed technique is to apply a hyperparameter-tuned deep learning method for identifying and classifying malware images. Primarily, the ShuffleNet method is mainly used to derivate the feature vectors. Besides, the snake optimization algorithm can be deployed to boost the choice of hyperparameters for the ShuffleNet algorithm. For the recognition and classification of malware images, attention-based bi-directional long short-term memory model. The simulation evaluation of the proposed algorithm has been examined using the Malimg malware dataset. The experimental values inferred that the proposed methodology achieves promising performance with a maximum accuracy of 98.42% compared to existing models.

Quality inspection and defect detection remain critical challenges across diverse industrial applications. Driven by advancements in Deep Learning, Convolutional Neural Networks (CNNs) have revolutionized Computer Vision, enabling breakthroughs in image analysis tasks like classification and object detection. CNNs’ feature learning and classification capabilities have made industrial defect detection through Machine Vision one of their most impactful applications. This article aims to showcase practical applications of CNN models for surface defect detection across various industrial scenarios, from pallet racks to display screens. The review explores object detection methodologies and suitable hardware platforms for deploying CNN-based architectures. The growing Industry 4.0 adoption necessitates enhancing quality inspection processes. The main results demonstrate CNNs’ efficacy in automating defect detection, achieving high accuracy and real-time performance across different surfaces. However, challenges like limited datasets, computational complexity, and domain-specific nuances require further research. Overall, this review acknowledges CNNs’ potential as a transformative technology for industrial vision applications, with practical implications ranging from quality control enhancement to cost reductions and process optimization.

Polytropic working conditions are prioritized by most intelligent adaptive methods. But the privacy and the inaccessibility of the source data during the transfer learning process are not considered. Moreover, the attention weights for each time point and neural network channel in the fault signals are also not addressed. Intending to deal with the problems above, we put forward a mixed attention network for source-free domain adaptation in bearing fault diagnosis work. We fully utilize the only-once source fault information to generate a source model with strong anomaly detection capabilities by our mixed attention network. Mixed attention network achieved an average accuracy of over 93% in both two datasets and achieved the highest accuracy in all tasks of the ablation experiment.

An electrocardiogram (ECG) is a non-invasive and cost-effective method for diagnosing heart disease. However, physicians often face challenges in interpreting ECG. As a result, deep learning (DL) models have been proposed to assist with interpretation. However, the development of a robust, interpretable model that performs well across diverse ECG datasets remains a focus. Hence, this study presents a robust interpretable DL-based system. The model utilizes a multi-channel hybrid architecture. It integrates 12 blocks of one-dimensional convolutional neural network (CNN) with bidirectional long short-term memory (BiLSTM) networks, followed by an attention mechanism and a two-dimensional CNN, and finally fully connected (FC) layers for classification. The model separately trained and tested on three 12-lead ECG datasets: PTB-XL, CODE-15%, and the reduced seven and merged four classes of Chapman Arrhythmia datasets, achieved average test accuracy rates of 89.84%, 97.82%, 98.55%, and 98.80%, respectively.The result indicates the model’s effectiveness across different ECG datasets. Besides, the classification output is analyzed using two post-hoc model interpretability techniques: Gradient-weighted Class Activation Mapping Plus Plus (Grad-CAM++) and SHapley Additive exPlanations (SHAP). These techniques are applied to the trained model to visualize influential segments of the ECG signal, both at the instance level for specific samples and at the test set level to assess the contributions of individual ECG leads among the 12 leads that influence the model predictions. The model’s performance and its output interpretation techniques makes it a practicable tool in ECG-based heart disease diagnosis.

Object recognition along with classification are necessary for many applications, such as surveillance systems, car plate recognition, traffic monitoring, and face detection. Unlike existing approaches, ours incorporates a wide range of important factors to improve recognition precision. The primary phase in the image accumulating process is preprocessing, when semantic segmentation proves its usefulness by accurately defining the physical borders of specific objects inside an image in addition to recognizing them. This paper presents a novel approach to accurate object recognition. Segmentation incorporates previously identified homologous and related groups after employing the K-means clustering technique to group analogous colors and spatial patterns. Convolutional Neural Network (CNN) technology is ultimately used to identify objects in different environmental circumstances. Performance metrics like as F1 Score=0.948, Precision = 0.968, and Recall=0.932 for MSRC and F1 Score=0.921, Precision = 0.951, and Recall=0.891 for Caltech 101 and F1 Score=0.847, Precision = 0.879, and Recall=0.827 over Pascal Voc 2012 demonstrate the efficiency of our strategy. The efficacy of the suggested method is evaluated using multiple benchmark datasets, MSRC-v2, Caltech 101 and Pascal Voc 2012, yielding recognition accuracies of 92.25%, 91.91% and 93.50% respectively, when tested against the Microsoft Research Cambridge (MSRC), California Institute of Technology 101 Object Categories (Caltech 101) and Pascal Voc 2012 datasets.

Recognizing intricate hand gestures is challenging in various applications like human-computer interaction and e-learning. Accurate classification is crucial for preserving cultural practices and enabling intuitive interaction with machines. However, existing models often struggle with challenges like dynamic lighting, complex backgrounds, varying camera angles, intricate hand poses, noise, and diverse hand attributes. To address the aforementioned challenges, we introduce EnGesto, an ensemble learning model for categorizing intricate hand gestures. EnGesto comprises three major modules: Data Augmentation (DAug), Extract Visual Geometry Group (EVGG), and Multistage Hand Gesture Classification (MuGest). DAug simulates real-world imaging conditions, enhancing accuracy in detecting hand movements and improving resilience to unexpected events, strengthening reliability and capacity. EVGG extracts feature maps using a customized VGG16 model. Within the MuGest module, advanced components such as Fully Convolutional Network (FCN), Region Proposal Network (RPN), Convolutional Neural Network (CNN), Global Max Pooling, Attention layer, and a Fully Connected (FC) layer are employed to carefully select relevant features from EVGG to achieve precise, robust hand gesture classification. Research showcased exemplary performance of the proposed model, surpassing its counterparts in classification accuracy of 97.85%, outperforms VGGNet, ResNet, EfficientNet, and CNN even under demanding image conditions, as in Indian classical dance, Bharatanatyam, with its core mudras—precise gestures conveying a range of emotions and ideas. EnGesto excels in accurate gesture classification, enhancing precision, and efficiency in preserving and facilitating e-learning of treasured art forms while promoting cultural significance, enabling natural, intuitive interaction with machines, and opening avenues for further research and development in this domain.

With the ongoing development of Indoor Location-Based Services, the location information of users in indoor environments has been a challenging issue in recent years. Due to the widespread use of WiFi networks, WiFi fingerprinting has become one of the most practical methods of locating mobile users. In addition to localization accuracy, some other critical factors such as latency, and users’ privacy should be considered in indoor localization systems. In this study, we propose a light Convolutional Neural Network-based method for edge devices (e.g. smartphones) to overcome the above issues by eliminating the need for a cloud/server in the localization system. The proposed method is evaluated for three different open datasets, i.e., UJIIndoorLoc, Tampere and UTSIndoorLoc, as well as for our collected dataset named SBUK-D to verify its scalability. We also evaluate performance efficiency of our localization method on an Android smartphone to demonstrate its applicability to edge devices. For UJIIndoorLoc dataset, our model obtains approximately 99% building accuracy, over 90% floor accuracy, and 9.5 m positioning mean error with the model size and inference time of 0.5 MB and $51~\mu $ s, respectively, which demonstrate high accuracy in range of state of the art works as well as amenability to the resource-constrained edge devices.

With the development of artificial intelligence technology, time series classification has attracted greater attention. Various methods have been considered for using deep learning models to perform the task. Training such models, however, requires a large amount of high-quality labeled samples, which may not be available due to the expensive cost of collection. Therefore, we proposed a novel framework named Convolutional Neural Network based Stepwise Improving Generative Adversarial Network (SIGAN-CNN) to solve this problem by generating more samples from the original distribution. We designed the structure of the generator and discriminator of SIGAN to be suitable for time series data. We also explored a method for extracting trend information from time series data to obtain trend samples for stepwise training. Therefore, the generator can fit the original time series data with subtle variations. More importantly, SIGAN can improve the diversity of the generated samples and the stability of the training process to achieve a higher quality of the generated samples. The generated samples are then combined with CNN based time series data classification methods to improve the classification performance of time series data. Especially, the combination of SIGAN and Multi-scale Attention Convolutional Neural Network (MACNN) is suitable for this task. We conducted a comprehensive evaluation of 8 standard datasets from various domains. The results demonstrate that SIGAN-MACNN achieves the best performance and outperforms the other state-of-the-art methods by a large margin. Therefore, SIGAN-MACNN offers an effective solution for addressing the time series classification task of small sample size.

Automatic and accurate classification of thyroid nodules is of great significance to doctors for clinical diagnosis and subsequent treatment recommendations. Since there are no obvious features between benign and malignant nodules, enlarging or reducing the image will result in blurred edges and image distortion, thus limiting the accuracy of clinical diagnosis. Furthermore, the prevalence of sample class imbalance in medical images poses significant challenges to applying convolutional neural networks in thyroid nodule classification methods. This paper proposes a network Tnc-Net for thyroid nodule classification. The network backbone can adapt to the problem of small data volume, capture global features with the help of simple channel attention, and effectively extract image information. The branch network supplements the feature extraction from the backbone network, and the information extracted from the backbone and branch networks is effectively utilized through the fusion module. In addition, this article designs training strategies suitable for this network to deal with category imbalance, improve model classification performance, and make classification results more clinically referenceable. The method test accuracy is 0.902, which exceeds other classic deep learning models in classification. This result demonstrates the effectiveness of our method in achieving automatic classification of thyroid nodules.

Sign language is the predominant mode of communication for individuals with auditory impairment. In Bangladesh, BdSL or Bangla Sign Language is widely used among the hearing-impaired population. However, because of the general public’s limited awareness of sign language, communicating with them using BdSL can be challenging. Consequently, there is a growing demand for an automated system capable of efficiently understanding BdSL. For automation, various Deep Learning (DL) architectures can be employed to translate Bangla Sign Language into readable digital text. The automation system incorporates live cameras that continuously capture images, which a DL model then processes. However, factors such as lighting, background noise, skin tone, hand orientations, and other aspects of the image circumstances may introduce uncertainty variables. To address this, we propose a procedure that reduces these uncertainties by considering three modalities: spatial information, skeleton awareness, and edge awareness. We introduce three image pre-processing techniques alongside three CNN models. The CNN models are combined using nine distinct ensemble meta-learning algorithms, with five of them being modifications of averaging and voting techniques. In the result analysis, our individual CNN models achieved higher training accuracy at 99.77%, 98.11%, and 99.30%, respectively, than most of the other state-of-the-art image classification architectures, except for ResNet50, which achieved 99.87%. Meanwhile, the ensemble model attained the highest accuracy of 95.13% on the testing set, outperforming all individual CNN models. This analysis demonstrates that considering multiple modalities can significantly improve the system’s overall performance in hand pattern recognition.

This research introduces a novel approach called Circular Mesh Network (CirMNet), a shape-based hybrid feature extraction technique. It also proposes innovative Fractal Dimension (FD) and statistical feature extraction techniques for vertically symmetrical images. Convolutional Neural Networks (CNNs) have gained widespread popularity across various domains, including the interpretation of medical images. CNNs excel at extracting prominent features in the initial layers and progressively learn to capture more complex features as they advance. However, the pooling and striding operations involved in CNNs can lead to a loss of spatial and structural details in the image because CNNs require a mechanism to preserve the internal representation and describe the intricate relationships between image components or pixels. Circular Mesh-based Shape and Margin Descriptor (CMSMD) focuses on extracting structural, statistical, and property-based features. However, it does not encompass features such as texture or color. The objective of CirMNet is to leverage the strengths of both CNNs and CMSMD, and to mitigate their respective weaknesses. Structural and texture features were generated from CirMNet, and its performance was evaluated for the diagnosis of neurodegenerative disorders, particularly Alzheimer’s Disease (AD). The model can easily identify the permanent shrinkage and destruction of brain cells in MRIs of patients and exhibited a notable accuracy of 97.34% in classifying various stages of AD, encompassing Control, Early Mild Cognitive Impairment (EMCI), and Late Mild Cognitive Impairment (LMCI). This achievement represents a substantial improvement over the existing state-of-the-art methods in the domain.

Aspect-based sentiment classification, as a more fine-grained sentiment analysis task, focuses on predicting the sentiment tendency expressed in a sentence based on specific aspects. However, current text sentiment analysis models face challenges when dealing with long comments posted by users on social media, as users often do not explicitly mention sentiment aspects directly in their comments. This paper focuses on aspect extraction and sentiment classification. By constructing a neural network model that integrates a self-attention mechanism, the model is able to learn word embeddings that incorporate contextual semantic information. Furthermore, the author introduce a self-attention mechanism based on relative position representations, which aims to simulate the order of words and achieve parallelized training of inputs by reducing parameters, while simultaneously extracting aspect and sentiment features. Additionally, the author designed a convolutional neural network model and utilized the ReLu gate to selectively output sentiment features based on the given aspect category, while implementing the K-max pooling technique. Comparative experiments conducted on three standard datasets, SemEval, Tweets, and CVAT, showed that this model achieved average best performance on all three datasets. Specifically, on the SemEval dataset, when predicting valence values, the MSE, MAE, and Pearson correlation coefficient reached optimal values of 1.00, 0.88, and 0.73, respectively. While the overall performance on the CVAT dataset was slightly lower, this model still achieved the best MSE of 0.89, MAE of 0.81, and Pearson correlation coefficient of 0.64 when predicting arousal values. This result demonstrates that this method provides relatively balanced and excellent performance in predicting both valence and arousal, validating its practical application value in the field of sentiment analysis.

Due to the high incidence of lumbar vertebral lesions causing spondylolisthesis and intervertebral disc protrusion, lumbar spine (vertebrae and intervertebral discs) MRI image segmentation can provide effective clinical information for the initial diagnosis and early treatment of current lumbar spine-related diseases. However, in MRI images, there is a significant overlap and similarity in features between the vertebral bones and intervertebral discs within the lumbar spine. Therefore, the effective identification and segmentation of each vertebra and intervertebral disc in the lumbar spine pose a significant challenge. We propose a lumbar spine MRI segmentation model based on the 3D Residual U-Net, incorporating boundary segmentation structures and a hybrid attention mechanism. The model achieves boundary-constrained segmentation of individual vertebrae and intervertebral discs by integrating the boundary segmentation module. Additionally, it utilizes a hybrid attention module based on convolutional attention and self-attention mechanisms for multiscale feature extraction in the lumbar spine. The proposed model is trained and validated using the publicly available datasets MRSpineSeg2021 and SpineSagT2Wdataset3. Experimental results demonstrate a significant improvement in segmentation performance, as measured by metrics such as the Dice similarity coefficient (DSC) and Hausdorff distance (HD). This validates the superiority and generalization performance of our proposed lumbar spine MRI.

For distribution networks, accurate short-term load forecasting is a prerequisite for the safe and stable operation as well as economically optimized dispatching of the grid. In order to enhance the accuracy of short-term power load forecasting, this paper proposes a forecasting method that combines convolutional neural network (CNN), long short-term memory (LSTM) network, and the Neuralprophet model. This method utilizes the Neuralprophet model to capture trends, seasonal cycles, holiday activities, and other components within load data, while leveraging the data feature extraction capability of the CNN model and the long-term sequence prediction ability of the LSTM model. The optimal hyperparameters of the models are determined using the Bayesian optimization algorithm, and the predictions of the two models are fused through the least squares method. Application of this method to forecasting on various load datasets demonstrates its superior prediction accuracy compared to other classical models.