Feature Extraction Network Research Articles

Intelligent detection of airport pavement subtle crack based on one-stage anchor-free deep neural network

ABSTRACT To solve issues with low-detection accuracy and high-missed detection rate in subtle pavement cracks, this research proposed an intelligent detection method based on a one-stage anchor-free deep neural network, which consists of a feature extraction network, detection neck, and detection head. The feature extraction network with an extended deep attention network is responsible for extracting the features and output at different scales. The detection neck containing the special spatial pyramid structure fuses the contextual information of different scales to improve the abstract features. The detection head is composed of the anchor-free mechanism with decoupling property and the reparameterized structure to improve the detection accuracy and balance the algorithm complexity. The experimental results demonstrate that the Precision, Recall, and F1-Score are, 92.74%, 91.92%, and 0.9233 respectively. It has advantages over the most advanced one-stage detection method. Based on testing results, it has prominent stability and robustness for subtle pavement crack detection.

Image retrievable encryption based on linear fitting and orthogonal transformation

Abstract With the development of cloud computing, an increasing number of resource-constrained image owners tend to store their images in the cloud and rely on image retrieval services to obtain the images they desire. However, the security of the cloud cannot be fully guaranteed. To ensure image security while achieving good retrieval performance, we have designed a retrievable image encryption algorithm based on linear fitting and orthogonal transformation. This algorithm first generates encryption and feature extraction domains through orthogonal decomposition, and then applies a modified ResNet50 network for feature extraction in the feature extraction domain. The encryption process employs an improved affine transformation based on linear fitting, where part of the fitting values comes from the original image data and the other part comes from data generated by a chaotic system. Additionally, to simplify the measurement of feature similarity in the cloud, we have designed a hierarchical feature index tree to narrow the retrieval scope, thereby reducing retrieval complexity. Experimental results show that the proposed algorithm effectively protects image privacy and achieves high retrieval accuracy. The F-score reached 6.7634% on the Ghim10k dataset and 25.514% on the Corel 1K dataset, significantly improving upon traditional methods. This algorithm has potential application value in the fields of secure image storage and efficient retrieval in the cloud.

Deep Learning Algorithm for Optimized Sensor Data Fusion in Fault Diagnosis and Tolerance

Environmental perception is one of the key technologies to realize autonomous vehicles. The fault diagnosis process involves identifying the fault that occurred or the cause of the out-of-control condition. Here, the major objective is to locate problems in detection by analysing previous data or sequential patterns of data that cause failure. This study evaluates the use of deep learning for improved sensor data fusion in fault identification and tolerance using the KITTI dataset. The input video from the dataset has been transformed to frames through median filtering. Next, feature extraction is applied to a preprocessed image, resulting in the fusion of sensor data. Data fusion is then carried out utilizing an enhanced RPN (region proposal network). The enhanced RPN also has a loss function (object detection loss, bounding box loss and target classification loss), an estimate of ROI and feature extraction network (FEN). Through the use of the COOT connected blue monkey optimization (CCBMO) model, the weight of the optimally enhanced RPN is established. Next, using global non-maximum suppression with both global and local confidence, fault identification and tolerance are carried out. From the analysis, it clearly shows that proposed method accomplished better results in terms of accuracy, precision and specificity of 97.78%, 93.76% and 93.43%, respectively, when compared with various conventional models with respect to diverse performance measures.

FastQAFPN-YOLOv8s-Based Method for Rapid and Lightweight Detection of Walnut Unseparated Material

Walnuts possess significant nutritional and economic value. Fast and accurate sorting of shells and kernels will enhance the efficiency of automated production. Therefore, we propose a FastQAFPN-YOLOv8s object detection network to achieve rapid and precise detection of unsorted materials. The method uses lightweight Pconv (Partial Convolution) operators to build the FasterNextBlock structure, which serves as the backbone feature extractor for the Fasternet feature extraction network. The ECIoU loss function, combining EIoU (Efficient-IoU) and CIoU (Complete-IoU), speeds up the adjustment of the prediction frame and the network regression. In the Neck section of the network, the QAFPN feature fusion extraction network is proposed to replace the PAN-FPN (Path Aggregation Network—Feature Pyramid Network) in YOLOv8s with a Rep-PAN structure based on the QARepNext reparameterization framework for feature fusion extraction to strike a balance between network performance and inference speed. To validate the method, we built a three-axis mobile sorting device and created a dataset of 3000 images of walnuts after shell removal for experiments. The results show that the improved network contains 6071008 parameters, a training time of 2.49 h, a model size of 12.3 MB, an mAP (Mean Average Precision) of 94.5%, and a frame rate of 52.1 FPS. Compared with the original model, the number of parameters decreased by 45.5%, with training time reduced by 32.7%, the model size shrunk by 45.3%, and frame rate improved by 40.8%. However, some accuracy is sacrificed due to the lightweight design, resulting in a 1.2% decrease in mAP. The network reduces the model size by 59.7 MB and 23.9 MB compared to YOLOv7 and YOLOv6, respectively, and improves the frame rate by 15.67 fps and 22.55 fps, respectively. The average confidence and mAP show minimal changes compared to YOLOv7 and improved by 4.2% and 2.4% compared to YOLOv6, respectively. The FastQAFPN-YOLOv8s detection method effectively reduces model size while maintaining recognition accuracy.

Breast histopathological imaging using ultra-fast fluorescence confocal microscopy to identify cancer lesions at early stage.

Ultrafast fluorescent confocal microscopy is a hypothetical approach for breast cancer detection because of its potential to achieve instantaneous, high-resolution images of cellular-level tissue features. Traditional approaches such as mammography and biopsy are laborious, invasive, and inefficient; confocal microscopy offers many benefits over these approaches. However, confocal microscopy enables the exact differentiation of malignant cells, the expeditious examination of extensive tissue sections, and the optical sectioning of tissue samples into tiny slices. The primary goal should be to prevent cancer altogether, although detecting it early can help achieve that objective. This research presents a novel Breast Histopathology Convolutional Neural Network (BHCNN) for feature extraction and recursive feature elimination method for selecting the most significant features. The proposed approach utilizes full slide images to identify tissue in regions affected by invasive ductal carcinoma. In addition, a transfer learning approach is employed to enhance the performance and accuracy of the models in detecting breast cancer, while also reducing computation time by modifying the final layer of the proposed model. The results showed that the BHCNN model outperformed other models in terms of accuracy, achieving a testing accuracy of 98.42% and a training accuracy of 99.94%. The confusion matrix results show that the IDC positive (+) class achieved 97.44% accuracy and 2.56% inaccurate results, while the IDC negative (-) class achieved 98.73% accuracy and 1.27% inaccurate results. Furthermore, the model achieved less than 0.05 validation loss. RESEARCH HIGHLIGHTS: The objective is to develop an innovative framework using ultra-fast fluorescence confocal microscopy, particularly for the challenging problem of breast cancer diagnosis. This framework will extract essential features from microscopy and employ a gradient recurrent unit for detection. The proposed research offers significant potential in enhancing medical imaging through the provision of a reliable and resilient system for precise diagnosis of breast cancer, thereby propelling the progression of state-of-the-art medical technology. The most suitable feature was determined using BHRFE optimization techniques after retrieving the features by proposed model. Finally, the features chosen are integrated into a proposed methodology, which is then classified using a GRU deep model. The aforementioned research has significant potential to improve medical imaging by providing a complex and reliable system for precise evaluation of breast cancer, hence advancing the development of cutting-edge medical technology.

Breaking dimensional barriers in hyperspectral target detection: Atrous convolution with Gramian Angular field representations

Hyperspectral images contain extensive spectral bands with rich spectral information that reflects object properties. Leveraging state-of-the-art deep learning techniques has proven to be effective in hyperspectral target detection. However, compared to two-dimensional matrix data, the one-dimensional nature of spectral sequence limits the information that can be extracted, posing a challenge for deep learning-based hyperspectral target detection methodologies. To address this issue, a novel hyperspectral target detection method employing atrous convolution with gramian angular field representations is proposed in this paper. This approach breaks the barrier between one-dimensional vector and two-dimensional matrix by gramian angular field, transforming the spectral sequences from one-dimensional vectors into two-dimensional matrices, enabling the exploration of multidimensional relationships within spectral band relations through an atrous convolution-based spectral feature extraction network. The proposed model transcends the traditional one-dimensional spectral target detection limitations, offering a new perspective for spectral-based hyperspectral target detection. Experimental results on four real-world hyperspectral datasets demonstrate that the proposed method significantly outperforms existing state-of-the-art methods in detection performance, showcasing its potential for advancing hyperspectral target detection.

Neutron-Image Super-Resolution based on Convolutional Neural Networks and Novel Parallel Transformer

Enhancing the resolution of neutron imaging can be achieved by upgrading the hardware components of the neutron imaging system, including employing more potent neutron source systems, utilizing high-resolution detectors, and implementing superior collimation techniques. However, such improvements are frequently constrained by challenges such as high cost and limited flexibility. In this paper, a flexible and efficient single-neutron-image super-resolution network is proposed for neutron image super-resolution. We propose a novel multi-scale feature extraction network, which can extract features deeply by fusing features of different types and levels. In order to learn more context information, a novel parallel Transformer is proposed to overcome the long dependency problem of sequences. A high-frequency information enhancement network is proposed to enhance the high-frequency information of neutron image in order to address the blur problem. Quantitative and qualitative analyses of benchmark datasets and real neutron images show that the proposed network performs better on Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) than bicubic interpolation and other methods, and reconstructs high-resolution neutron images that are more consistent with human visual perception, highlighting its application potential in neutron images quality enhancement.

Transformer-guided Feature Pyramid Network for multi-view stereo

Feature Pyramid Network (FPN) is widely used in Multi-View Stereo (MVS) to extract multi-scale features, effectively enhancing both the quality and efficiency of reconstruction. However, the multi-scale features are simply fused by element-wise addition, ignoring the contextual relationship between them. To solve this problem, a Transformer-guided Feature Pyramid Network (TFPN-MVSNet) is proposed, which uses the self-attention mechanism to aggregate the long-range context information between multi-scale features. Firstly, the feature enhancement module is designed based on internal attention mechanism to calculate the importance of different parts within the feature sequence, which pays more attention to the features related to reconstruction. The feature aggregation module is designed based on cross attention mechanism to focus on the contextual information of multi-scale features. Secondly, the two modules are introduced into the multi-scale feature extraction network to enhance the three-stage features from coarse to fine. Finally, the Spatial Weighted Cost Volume Aggregation method (SWCA)uses the spatial attention mechanism to calculate the pixel weights at the same position in different views, which suppresses the negative impact of occluded pixels. Evaluated on the DTU dataset, the overall error is reduced by 0.062 mm compared to CasMVSNet, and by 0.012 mm and 0.019 mm compared to similar methods TransMVSNet and CostFormer, respectively. Our method also achieves competitive results on the Tanks & Temples and ETH3D datasets.

Advancements in Arabic Sign Language Recognition: A Method based on Deep Learning to Improve Communication Access

This research anticipates the issue of real-time sign language recognition as a means of facilitating the daily communication of the hearing-impaired persons. Improvement in the rate of real time sign language recognition will help in boosting interaction in the society and access to services for the hearing-impaired persons. This paper outlines a step-by-step guide on fine tuning an object detection model for the AASL using the SSD architecture. The AASL dataset which includes total of 7,857 samples of more than 200 patients are first preprocessed where the images are scaled, normalized, and augmented. In SSD model, the base of the network is the VGG16 network and few extra layers for feature extraction and for auxiliary and prediction of objects are added. Particularly in object detection, Intersection over Union (IoU), and mean Average Precision (mAP) are used in addition to the confusion matrix metrics comprising accuracy, precision, recall, and F1-score. The proposed model provides good recognition accuracy of 98 percent. This feature checks its capability of real-time identification of sign languages and shows 25% efficiency. The efficiency of the proposed method is explained when comparing the results of the modified VGG16-based SSD with other methodologies. As for the improvement of communication for the hearing-impaired individuals, the present work demonstrates that the deep learning methods can be significantly effective; however, suggesting that more research efforts should be directed to real-time solutions and that the datasets should be expanded.

CGBi_YOLO: Lightweight Land Target Detection Network

Object detection algorithms for optical remote sensing images often face challenges in computational efficiency, particularly when detecting small and densely packed targets. This paper introduces CGBi_YOLO, a novel lightweight land target detection network designed to optimize computational resource utilization while maintaining detection capabilities for small-scale targets. Our approach incorporates an innovative lightweight optimization strategy featuring a new lightweight backbone feature extraction network: CSPGhostNet. This model significantly enhances the detection ability of small objects within optical remote sensing images without increasing computational demands. The efficacy of the proposed model is validated through rigorous experimentation on the DOTA dataset. Compared to the baseline model, CGBi_YOLO achieves a 30% reduction in parameters and a 36% increase in inference speed. The model demonstrates exceptional performance in handling small and densely packed targets within optical remote sensing images, showcasing its potential for real-world applications in fields such as environmental monitoring, urban planning, and disaster management.

GPTrans: A Biological Language Model-Based Approach for Predicting Disease-Associated Mutations in G Protein-Coupled Receptors.

Accurately predicting mutations in G protein-coupled receptors (GPCRs) is critical for advancing disease diagnosis and drug discovery. In response to this imperative, GPTrans has emerged as a highly accurate predictor of disease-related mutations in GPCRs. The core innovation of GPTrans resides in the design of a novel feature extraction network, that is capable of integrating features from both wildtype and mutant protein variant sites, utilizing multifeature connections within a transformer framework to ensure comprehensive feature extraction. A key aspect of GPTrans's effectiveness is our introduction of an innovative deep feature integration strategy, which merges embeddings and class tokens from multiple protein language models, including evolutionary scale modeling and ProtTrans, thus shedding light on the biochemical properties of proteins. Leveraging transformer components and a self-attention mechanism, GPTrans captures higher-level representations of protein features. Employing both wildtype and mutation site information for feature fusion not only enriches the predictive feature set but also avoids the common issue of overestimation associated with sequence-based predictions. This approach distinguishes GPTrans, enabling it to significantly outperform existing methods. Our evaluations across diverse GPCR data sets, including ClinVar and MutHTP, demonstrate GPTrans's superior performance, with average AUC values of 0.874 and 0.590 in 10-fold cross-validation. Notably, compared to the AlphaMissense method, GPTrans exhibited a remarkable 38.03% improvement in accuracy when predicting disease-associated mutations in the MutHTP data set. A thorough analysis of the predicted results further validates the model's effectiveness. The source code, data sets, and prediction results for GPTrans are available for academic use at https://github.com/EduardWang/GPTrans.

An Adaptive Multimodal Fusion 3D Object Detection Algorithm for Unmanned Systems in Adverse Weather

Unmanned systems encounter challenging weather conditions during obstacle removal tasks. Researching stable, real-time, and accurate environmental perception methods under such conditions is crucial. Cameras and LiDAR sensors provide different and complementary data. However, the integration of disparate data presents challenges such as feature mismatches and the fusion of sparse and dense information, which can degrade algorithmic performance. Adverse weather conditions, like rain and snow, introduce noise that further reduces perception accuracy. To address these issues, we propose a novel weather-adaptive bird’s-eye view multi-level co-attention fusion 3D object detection algorithm (BEV-MCAF). This algorithm employs an improved feature extraction network to obtain more effective features. A multimodal feature fusion module has been constructed with BEV image feature generation and a co-attention mechanism for better fusion effects. A multi-scale multimodal joint domain adversarial network (M2-DANet) is proposed to enhance adaptability to adverse weather conditions. The efficacy of BEV-MCAF has been validated on both the nuScenes and Ithaca365 datasets, confirming its robustness and good generalization capability in a variety of bad weather conditions. The findings indicate that our proposed algorithm performs better than the benchmark, showing improved adaptability to harsh weather conditions and enhancing the robustness of UVs, ensuring reliable perception under challenging conditions.

MSDFEN: Multi-scale dynamic feature extraction network for pathological voice detection

With the rapid development of deep learning methods, their application in pathological voice detection has become increasingly extensive, yielding promising results. However, most deep learning methods used in pathological voice detection employ network architectures where all constituent modules are static, performing the same operations on all inputs. This significantly limits the model’s adaptive capacity and generalization ability and restricts further improvement in model performance. To address this issue, this paper proposes a novel pathological voice detection system called the Multi-Scale Dynamic Feature Extraction Network (MSDFEN), designed to enhance the performance and adaptive capability of pathological voice detection systems. In the MSDFEN model, sinc filter banks combined with a channel attention mechanism were employed for the preprocessing of vocal signals, effectively capturing the high-frequency transitions characteristic of pathological voices. Furthermore, dynamic blocks, consisting of multiple dynamic components, were designed and integrated into a multi-scale convolutional neural network, significantly enhancing the network’s dynamic performance and enriching the features obtained through multi-scale fusion. Comparative experiments and ablation studies were conducted using three internationally recognized pathological voice detection databases: MEEI, SVD, and HUPA. the proposed model achieved recognition accuracies of 0.9883, 0.7424, and 0.8409 in these databases, and other parameters also yielded satisfactory results. Experimental results indicate that the proposed method exhibits excellent adaptive and generalization capabilities in pathological voice detection.

ICGM-YOLO: UAV remote sensing small target detection algorithm based on feature fusion enhancement module and ghost convolution

Abstract Remote sensing images captured at high altitudes are affected by background noise and low resolution, leading to suboptimal performance of commonly used small object detection algorithms. This paper introduces ICGM-YOLO, an enhanced approach designed to improve the precision of detecting small objects within remote sensing images. The methodology refines YOLOv8 for optimal performance in this context. First, within the feature extraction network, iterative attention feature fusion combines feature maps and integrates features from different layers. Second, the convolutional layer and C2f module are replaced with the GhostConv module and C3Ghost to extract features. Finally, this paper replaces CIoU with the MFIoU loss function, which combines the new bounding box MPDIoU based on minimum point distance with Focaler-IoU. This replacement accelerates model convergence and enhances detection recall. Experimental results from the Visdrone2019 dataset indicate that ICGM-YOLO outperforms the original YOLOv8 by achieving an 8.1% improvement in mAP0.5 detection rate and a 5.3% increase in mAP0.5:0.95 detection rate. Additionally, ICGM-YOLO reduces the parameter count by 48.2% and decreases computational complexity by 45.2%.

The research on dance motion quality evaluation based on spatiotemporal convolutional neural networks

With the rapid development of artificial intelligence and computer vision technology, dance action recognition and quality assessment have become a challenging and important research field. Traditional dance movement recognition methods often rely on manual feature extraction and expert scoring, which have problems such as strong subjectivity, low efficiency, and difficulty in handling complex movements. To address these issues, this study proposes a dance action recognition and quality assessment method based on Spatio Temporal Convolutional Neural Network (ST-CNN). Specifically, it includes a feature extraction network based on 3D spatiotemporal description operators and a multi-scale aggregated long short memory network for extracting spatial features of single frame images and fusing video sequence features composed of all frames. This approach can effectively capture dynamic changes in dance movements by combining information from spatial and temporal dimensions, thereby achieving more accurate recognition and evaluation. In addition, to address the high storage and computational requirements of convolutional neural networks during training, we integrated dance video sequences into an IoT architecture to address parallel recognition and evaluation of multiple videos. Finally, we validated the proposed method on the public datasets NTU-RGB 60 and NTU-RGB 120, and the experimental results demonstrated the effectiveness of the method.

Domain-Invariant Few-Shot Contrastive Learning for Hyperspectral Image Classification

In Hyperspectral Image (HSI) classification, acquiring large quantities of high-quality labeled samples is typically costly and impractical. Traditional deep learning methods are limited in such scenarios due to their dependence on sample quantities. To address this challenge, researchers have turned to Few-Shot Learning (FSL). Although existing FSL methods improve classification performance by enhancing domain invariance through domain adaptation, they often overlook the critical issue of high inter-class similarity and large intra-class variability. Moreover, during domain alignment, features of different categories within the same domain may become confused. To address these issues, this paper proposes a novel Domain-Invariant Few-Shot Contrastive Learning (DIFSCL) method, which combines domain adaptation and contrastive learning strategies to not only learn domain-invariant features but also significantly enhance inter-class discriminability. Based on this, we further design a multi-scale adaptive attention mechanism for a hyperspectral feature extraction network to more effectively extract and optimize generalized features within the DIFSCL framework, significantly improving intra-class consistency and inter-class discriminative capability of features. Experimental results on three widely used HSI datasets demonstrate that our method significantly outperforms existing techniques in few-shot classification tasks.

A Study on Filter-Based Adversarial Image Classification Models

In view of the fact that adversarial examples can lead to high-confidence erroneous outputs of deep neural networks, this study aims to improve the safety of deep neural networks by distinguishing adversarial examples. A classification model based on filter residual network structure is used to accurately classify adversarial examples. The filter-based classification model includes residual network feature extraction and classification modules, which are iteratively optimized by an adversarial training strategy. Three mainstream adversarial attack methods are improved, and adversarial samples are generated on the Mini-ImageNet dataset. Subsequently, these samples are used to attack the EfficientNet and the filter-based classification model respectively, and the attack effects are compared. Experimental results show that the filter-based classification model has high classification accuracy when dealing with Mini-ImageNet adversarial examples. Adversarial training can effectively enhance the robustness of deep neural network models.

Multimodal Brain Tumor Segmentation Based on Multi-Scale Feature Extraction Network

Brain tumors are prevalent and highly fatal conditions, making their precise detection crucial for effective treatment. Accurate brain tumor segmentation in magnetic resonance imaging (MRI) scans is often crucial for clinical diagnosis and developing more precise treatment strategies. MRI, a standard diagnostic tool for brain tumor identification, offers multiple modalities, each providing unique imaging characteristics. However, existing deep learning-based segmentation methods often struggle with varying tumor size, which can impact their effectiveness. To address these issues, we incorporate DeepLabv3+ into the task of brain tumor segmentation, leveraging its Atrous convolutional capabilities to capture comprehensive multi-scale features. By combining the multi-scale feature extraction strengths of DeepLabV3+ with the diverse imaging modalities of MRI, our approach mitigates the limitations associated with single-scale analysis, thereby enhancing diagnostic accuracy and supporting improved therapeutic outcomes. This article finds that this model achieves an accuracy of approximately 93.5% on the BraTS_2020 dataset.

Mine underground object detection algorithm based on TTFNet and anchor-free

Abstract To solve the problem of poor object detection effect caused by uneven light and high noise in underground mines, this study proposes a TTFNet (training-time-friendly network)-based object detection algorithm for underground mines. First, CenterNet and TTFNet algorithms are introduced, then pooling is introduced into CSPNet basic structure to design a lightweight feature extraction network, at the same time optimizing the feature fusion way in the original algorithm, optimizing residual shrinkage network structure, and introducing it into object detection task. Experiments were conducted on the established underground data set. The results show that compared with the original algorithm, our proposed algorithm can still maintain similar accuracy while significantly reducing model parameters; compared with other anchor-based detection algorithms, it has achieved similar overall performance.

An enhanced YOLOv8‐based bolt detection algorithm for transmission line

AbstractThe current bolt detection for overhead work robots used for transmission lines faces the problems of lightweight algorithms and high accuracy of target detection. To address these challenges, this paper proposes a lightweight bolt detection algorithm based on improved YOLOv8 (you only look once v8) model. Firstly, the C2f module in the feature extraction network is integrated with the self‐calibrated convolution module, and the model is streamlined by reducing spatial and channel redundancies of the network through the SRU and CUR mechanisms in the module. Secondly, the P2 small object detection layer is introduced into the neck structure and the BiFPN network structure is incorporated to enhance the bidirectional connection paths, thereby promoting the upward and downward propagation of features. It improves the accuracy of the network for bolt‐small target detection. The experimental results show that, compared to the original YOLOv8 model, the proposed algorithm demonstrates superior performance on a self‐collected dataset. The mAP accuracy is improved in this paper by 9.9%, while the number of model parameters and the model size is reduced by 0.973 × 106 and 1.7 MB, respectively. The improved algorithm improves the accuracy of the bolt detection while reducing the computation complexity to achieve more lightweight model.