Extract High-level Features Research Articles

Deep learning powered rapid lifetime classification of lithium-ion batteries

Lithium-ion batteries (LIBs) are currently the primary energy storage devices for modern electric vehicles (EVs). Early-cycle lifetime/quality classification of LIBs is a promising technology for many EV-related applications, such as fast-charging optimization design, production evaluation, battery pack design, second-life recycling, etc. The key challenge of the research problem is to develop an accurate classification method based on very limited early-cycle data, which contain very little information regarding battery degradation. To respond to such emerging need and tackle such technical challenge, this study develops a novel deep learning powered method for enabling the rapid LIB lifetime classification via very limited early-cycle data. First, the proposed method considers an innovative high-dimensional tensor input integrating early-cycle battery voltage, current, and temperature data to organically fuse the spatial, temporal, and physical battery information. Next, a convolutional sparse autoencoder-based feature engineering framework is developed to process such tensor input, automatically extract high-level latent features, and embed high-dimensional input information into a more compact representation. Finally, a regularized logistic regression model is developed to classify batteries into different lifetime groups based on a joint consideration of latent features as well as battery nominal and operational parameters. The effectiveness and robustness of the proposed method is verified on experimental data of battery degradation with three different chemistries and under multiple charge/discharge conditions. The performance of the proposed method is competitive by comparing with a set of well-known and recent benchmarking methods. In scenarios with only first-20-cycle degradation data available, the classification accuracy of the proposed method can reach 96.6%. In scenarios with only first-5-cycle data available, our classification accuracy can still reach 92.1%.

Hierarchical deep learning approach using fusion layer for Source Camera Model Identification based on video taken by smartphone

Over the last decade, videos uploaded and shared through web-based multimedia platforms and mobile applications have proliferated worldwide. This is because cloud-based applications such as iCloud, YouTube, Facebook, Twitter, and WhatsApp offer affordable and secure environments for video storage and sharing. However, new challenges have emerged alarming forensic analysts and investigators since videos can be used to commit heinous crimes such as blackmail, fraud, and forgery. Source Camera Identification (SCI) has become of paramount importance in the field of image and video forensics. Camera model identification can also help identify the perpetrators or narrow down the search and can be used to enhance SCI systems. In this context, existing approaches such as the Photo Response Non-Uniformity (PRNU) based methods and machine learning techniques such as the support vector machine (SVM) and deep learning models are commonly used solutions. This work exploits these two categories of methods by exploring a hierarchical deep learning model for camera model identification based on smartphone videos. The PRNU features are extracted by CNN-based structures during the training process. Proposed six-stream networks are leveraged to extract both low-level and high-level features through the network. A fusion layer is created based on joint sparse representation using forward and backward functions defined for fusing the proposed six streams. The proposed approach has been implemented and evaluated through intensive experiments, and results showed successful camera model identification with a performance at the frame level reaching an average accuracy of 69.9% for the Daxing dataset and 81.6% for the QUFVD dataset.

Impact of Deep Convolutional Neural Network Structure on Photovoltaic Array Extraction from High Spatial Resolution Remote Sensing Images

Accurate information on the location, shape, and size of photovoltaic (PV) arrays is essential for optimal power system planning and energy system development. In this study, we explore the potential of deep convolutional neural networks (DCNNs) for extracting PV arrays from high spatial resolution remote sensing (HSRRS) images. While previous research has mainly focused on the application of DCNNs, little attention has been paid to investigating the influence of different DCNN structures on the accuracy of PV array extraction. To address this gap, we compare the performance of seven popular DCNNs—AlexNet, VGG16, ResNet50, ResNeXt50, Xception, DenseNet121, and EfficientNetB6—based on a PV array dataset containing 2072 images of 1024 × 1024 size. We evaluate their intersection over union (IoU) values and highlight four DCNNs (EfficientNetB6, Xception, ResNeXt50, and VGG16) that consistently achieve IoU values above 94%. Furthermore, through analyzing the difference in the structure and features of these four DCNNs, we identify structural factors that contribute to the extraction of low-level spatial features (LFs) and high-level semantic features (HFs) of PV arrays. We find that the first feature extraction block without downsampling enhances the LFs’ extraction capability of the DCNNs, resulting in an increase in IoU values of approximately 0.25%. In addition, the use of separable convolution and attention mechanisms plays a crucial role in improving the HFs’ extraction, resulting in a 0.7% and 0.4% increase in IoU values, respectively. Overall, our study provides valuable insights into the impact of DCNN structures on the extraction of PV arrays from HSRRS images. These findings have significant implications for the selection of appropriate DCNNs and the design of robust DCNNs tailored for the accurate and efficient extraction of PV arrays.

Sentiment Analysis using a CNN-BiLSTM Deep Model Based on Attention Classification

With the rapid development of the Internet, the number of social media and e-commerce platforms increased dramatically. Users from all over world share their comments and sentiments on the Internet become a new tradition. Applying natural language processing technology to analyze the text on the Internet for mining the emotional tendencies has become the main way in the social public opinion monitoring and the after-sale feedback of manufactory. Thus, the study on text sentiment analysis has shown important social significance and commercial value. Sentiment analysis is a hot research topic in the field of natural language processing and data mining in recent ten years. The paper starts with the topic of "Sentiment Analysis using a CNN-BiLSTM deep model based on attention mechanism classification". First, it conducts an in-depth investigation on the current research status and commonly used algorithms at home and abroad, and briefly introduces and analyzes the current mainstream sentiment analysis methods. As a direction of machine learning, deep learning has become a hot research topic in emotion classification in the field of natural language processing. This paper uses deep learning models to study the sentiment classification problem of short and long text sentiment classification tasks. The main research contents are as follows. Firstly, Traditional neural network based short text classification algorithms for sentiment classification is easy to find the errors. The feature dimension is too high, and the feature information of the pool layer is lost, which leads to the loss of the details of the emotion vocabulary. To solve this problem, the Word Vector Model (Word2vec), Bidirectional Long-term and Short-term Memory networks (BiLSTM) and convolutional neural network (CNN) are combined in Quora dataset. The experiment shows that the accuracy of CNN-BiLSTM model associated with Word2vec word embedding achieved 91.48%. This proves that the hybrid network model performs better than the single structure neural network in short text. Convolutional neural network (CNN) models use convolutional layers and maximum pooling or max-overtime pooling layers to extract higher-level features, while LSTM models can capture long- term dependencies between words hence are better used for text classification. However, even with the hybrid approach that leverages the powers of these two deep-learning models, the number of features to remember for classification remains huge, hence hindering the training process. Secondly, we propose an attention based CNN-BiLSTM hybrid model that capitalize on the advantages of LSTM and CNN with an additional attention mechanism in IMDB movie reviews dataset. In the experiment, under the control of single variable of Data volume and Epoch, the proposed hybrid model was compared with the results of various indicators including recall, precision, F1 score and accuracy of CNN, LSTM and CNN-LSTM in long text. When the data size was 13 k, the proposed model had the highest accuracy at 0.908, and the F1 score also showed the highest performance at 0.883. When the epoch value for obtaining the optimal accuracy of each model was 10 for CNN, 14 for LSTM, 5 for MLP and 15 epochs for CNN-LSTM, which took the longest learning time. The F1 score also showed the best performance of the proposed model at 0.906, and accuracy of the proposed model was the highest at 0.929. Finally, the experimental results show that the bidirectional long- and short-term memory convolutional neural network (BiLSTM-CNN) model based on attention mechanism can effectively improve the performance of sentiment classification of data sets when processing long-text sentiment classification tasks. Keywords: sentiment analysis, CNN, BiLSTM, attention mechanism, text classification

Time-Series Neural Network: A High-Accuracy Time-Series Forecasting Method Based on Kernel Filter and Time Attention

This research introduces a novel high-accuracy time-series forecasting method, namely the Time Neural Network (TNN), which is based on a kernel filter and time attention mechanism. Taking into account the complex characteristics of time-series data, such as non-linearity, high dimensionality, and long-term dependence, the TNN model is designed and implemented. The key innovations of the TNN model lie in the incorporation of the time attention mechanism and kernel filter, allowing the model to allocate different weights to features at each time point, and extract high-level features from the time-series data, thereby improving the model’s predictive accuracy. Additionally, an adaptive weight generator is integrated into the model, enabling the model to automatically adjust weights based on input features. Mainstream time-series forecasting models such as Recurrent Neural Networks (RNNs) and Long Short-Term Memory Networks (LSTM) are employed as baseline models and comprehensive comparative experiments are conducted. The results indicate that the TNN model significantly outperforms the baseline models in both long-term and short-term prediction tasks. Specifically, the RMSE, MAE, and R2 reach 0.05, 0.23, and 0.95, respectively. Remarkably, even for complex time-series data that contain a large amount of noise, the TNN model still maintains a high prediction accuracy.

Fully automated sinogram-based deep learning model for detection and classification of intracranial hemorrhage.

To propose an automated approach for detecting and classifying Intracranial Hemorrhages (ICH) directly from sinograms using a deep learning framework. This method is proposed to overcome the limitations of the conventional diagnosis by eliminating the time-consuming reconstruction step and minimizing the potential noise and artifacts that can occur during the Computed Tomography (CT) reconstructionprocess. This study proposes a two-stage automated approach for detecting and classifying ICH from sinograms using a deep learning framework. The first stage of the framework is Intensity Transformed Sinogram Sythesizer, which synthesizes sinograms that are equivalent to the intensity-transformed CT images. The second stage comprises of a cascaded Convolutional Neural Network-Recurrent Neural Network (CNN-RNN) model that detects and classifies hemorrhages from the synthesized sinograms. The CNN module extracts high-level features from each input sinogram, while the RNN module provides spatial correlation of the neighborhood regions in the sinograms. The proposed method was evaluated on a publicly available RSNA dataset consisting of a large sample size of 8652patients. The results showed that the proposed method had a notable improvement as high as 27% in patient-wise accuracies when compared to state-of-the-art methods like ResNext-101, Inception-v3 and Vision Transformer. Furthermore, the sinogram-based approach was found to be more robust to noise and offset errors in comparison to CT image-based approaches. The proposed model was also subjected to a multi-label classification analysis to determine the hemorrhage type from a given sinogram. The learning patterns of the proposed model were also examined for explainability using the activationmaps. The proposed sinogram-based approach can provide an accurate and efficient diagnosis of ICH without the need for the time-consuming reconstruction step and can potentially overcome the limitations of CT image-based approaches. The results show promising outcomes for the use of sinogram-based approaches in detecting hemorrhages, and further research can explore the potential of this approach in clinicalsettings.

EEG Emotion Recognition by Fusion of Multi-Scale Features.

Electroencephalogram (EEG) signals exhibit low amplitude, complex background noise, randomness, and significant inter-individual differences, which pose challenges in extracting sufficient features and can lead to information loss during the mapping process from low-dimensional feature matrices to high-dimensional ones in emotion recognition algorithms. In this paper, we propose a Multi-scale Deformable Convolutional Interacting Attention Network based on Residual Network (MDCNAResnet) for EEG-based emotion recognition. Firstly, we extract differential entropy features from different channels of EEG signals and construct a three-dimensional feature matrix based on the relative positions of electrode channels. Secondly, we utilize deformable convolution (DCN) to extract high-level abstract features by replacing standard convolution with deformable convolution, enhancing the modeling capability of the convolutional neural network for irregular targets. Then, we develop the Bottom-Up Feature Pyramid Network (BU-FPN) to extract multi-scale data features, enabling complementary information from different levels in the neural network, while optimizing the feature extraction process using Efficient Channel Attention (ECANet). Finally, we combine the MDCNAResnet with a Bidirectional Gated Recurrent Unit (BiGRU) to further capture the contextual semantic information of EEG signals. Experimental results on the DEAP dataset demonstrate the effectiveness of our approach, achieving accuracies of 98.63% and 98.89% for Valence and Arousal dimensions, respectively.

Partial discharge based recognition of water droplets location in high voltage insulator using convolutional neural network – Bacterial foraging algorithm based optimized machine learning classifier

Measurement and analysis of Partial Discharge (PD) patterns have appeared as an emerging field in assessing insulation failure in High Voltage apparatus. This paper uses a PD signal combined with the deep convolution-optimized learning machine classifier (DC-OLMC) to predict the location of water droplets in 11 kV polymer insulators subjected to alternating currents. There are two major confront when applying the proposed algorithm: i) Contamination is a significant issue in PD signal measurement, which causes a reduction in recognition rate (RR), and ii) with minimal computing time, high-level feature extraction and recognition. Traditional condition monitoring methods of insulators concentrated on extracting fewer priority features from the input patterns. In the current work, to address this problem, an Alexnet with Bacterial Foraging Algorithm (BFO) based optimized kernel parameter classifier and Translation Invariant Wavelet Transform (TIWT) is employed to remove interference from PD signals. The analysis demonstrates that the suggested technique, with an identification rate of 99.17%, is considered a valuable tool for locating water droplets in high-voltage insulators.

Automatic borehole fracture detection and characterization with tailored Faster R-CNN and simplified Hough transform

Understanding the geological information of fractures from borehole images is essential for conducting a thorough geological analysis in engineering. Recently, the convolutional neural network (CNN) has been employed to extract high-level features of fractures from images. We propose an improved Faster R-CNN (region-based CNN) model for fracture detection, which is suitably equipped with a tailored region proposal network f-RPN for borehole fractures. The anchors generated by f-RPN are specifically designed with a fixed width and varying heights, and only the anchors with the maximum intersection-over-union values are labeled as positive. By taking advantage of the cyclic feature of borehole images, the images are expanded in four orientations to increase the amount and diversity of the model input. The detection results of the model can further be utilized to simplify the Hough transform for characterizing dip angles and dip directions of fractures. The proposed method is validated with borehole images from multiple projects in hydraulic engineering. Experiments reveal that the average detection accuracy is approximately 91.5%, which exhibits an improvement of 11.06% compared to the original Faster R-CNN. Further analysis indicates that both f-RPN and the multi-orientation expansion contribute to improving the prediction accuracy. It is also shown that this method demonstrates robustness against blurred image background and outperforms traditional image processing methods in the presence of noise. The fracture characterization by the simplified Hough transform turns out to be computationally efficient and accurate, with the absolute errors less than 5° compared to the manual identification results.

PSNet: Parallel-Convolution-Based U-Net for Crack Detection with Self-Gated Attention Block

Crack detection is an important task for road maintenance. Currently, convolutional neural-network-based segmentation models with attention blocks have achieved promising results, for the reason that these models can avoid the interference of lights and shadows. However, by carefully examining the structure of these models, we found that these segmentation models usually use down-sampling operations to extract high-level features. This operation reduces the resolution of features and causes feature information loss. Thus, in our proposed method, a Parallel Convolution Module (PCM) was designed to avoid feature information loss caused by down-sampling. In addition, the attention blocks in these models only focused on selecting channel features or spatial features, without controlling feature information flow. To solve the problem, a Self-Gated Attention Block (SGAB) was used to control the feature information flow in the attention block. Therefore, based on the ideas above, a PSNet with a PCM and SGAB was proposed by us. Additionally, as there were few public datasets for detailed evaluation of our method, we collected a large dataset by ourselves, which we named the OAD_CRACK dataset. Compared with the state-of-the-art crack detection method, our proposed PSNet demonstrated competitive segmentation performance. The experimental results showed that our PSNet could achieve accuracies of 92.6%, 81.2%, 98.5%, and 76.2% against the Cracktree200, CRACK500, CFD, and OAD_CRACK datasets, respectively, which were 2.6%, 4.2%, 1.2%, and 3.3% higher than those of the traditional attention models.

ESA-Net: An efficient scale-aware network for small crop pest detection

Pest detection aims to locate and classify the pests that may be present in the image, which plays a crucial role in the early warning of pests linked to the agriculture industry. However, crop pests are generally very small and uneven in scale, resulting in poor detection results. To address this challenge, we propose the Efficient Scale-Aware Network (ESA-Net), which comprises three core components: a High-Level Semantic Feature Extraction Module (HSFEM), a Low-Level Feature Enhancement Module (LFEM), and a Dynamic Scale-Aware Head (DSH). To avoid network optimization confusion, we use the HSFEM to extract high-level features that build the feature pyramid networks. Then, the LFEM is designed to further optimize the integrated feature map of low-level features, providing more fine-gained information for feature extraction and improving the detector's ability to differentiate between pests with similar appearances. Additionally, the proposed DSH can improve the detection performance of small objects by adaptively selecting the appropriate detection receptive field in accordance with various object scales. Experiment results on LMPD2020 dataset (68.8% mAP) and APHIDc dataset (75.3% mAP) demonstrate that ESA-Net achieves competitive results compared to the state-of-the-art methods.

Heterogeneous Convolutional Neural Networks for Emotion Recognition Combined with Multimodal Factorised Bilinear Pooling and Mobile Application Recommendation

The field of emotion recognition has garnered considerable interest due to its diverse applications in mental health, personalised advertising and enhancing user experiences. This research paper introduces a unique and innovative method for emotion recognition by integrating heterogeneous convolutional neural networks (CNNs) with multimodal factorised bilinear pooling. Furthermore, the paper also incorporates the integration of mobile application recommendations as part of the overall approach. The proposed method leverages the power of CNNs to extract high-level features from different modalities, including facial expressions, speech signals and physiological signals. By using heterogeneous CNNs, each modality is processed independently to capture modality-specific emotional cues effectively. To fuse the extracted features, multimodal factorised bilinear pooling is employed, which captures the complex interactions between different modalities while reducing the computational complexity. This pooling technique efficiently combines the modality-specific features, resulting in a compact and discriminative representation of the emotional state. In addition to emotion recognition, this paper also introduces the integration of mobile app recommendations. By leveraging the recognised emotion, the system recommends relevant mobile applications that are tailored to the user’s emotional state. This integration enhances user experience and facilitates emotion regulation through the utilisation of appropriate mobile apps. Experimental evaluations are conducted on benchmark emotion recognition datasets, including the DEAP and MAHNOB_HCI datasets. The findings of the study highlight the effectiveness of the proposed methodology in terms of accuracy and robustness, surpassing existing approaches in the field. Additionally, the integration of the mobile app recommendation system showcases encouraging outcomes by offering personalised recommendations tailored to the user’s emotional state.

Image Encryption Algorithm Combining Chaotic Image Encryption and Convolutional Neural Network

With the rapid development of information technology, the security of images has emerged as a significant area of research. This study presents an algorithm that integrates chaotic image encryption and a convolutional neural network (CNN) to enhance security and efficiency. The algorithm applies the properties of randomness and nonlinear mapping of chaotic sequences with the advanced feature extraction capabilities of a CNN model to achieve robust image encryption. First, we outline the fundamentals of chaotic image encryption and CNN. Chaotic image encryption employs chaotic sequence generation and nonlinear mapping to scramble pixel values for encryption purposes, while a CNN, as a deep-learning model with a local perceptual field and weight sharing, effectively extracts high-level image features. Subsequently, we provide a detailed description of the specific steps involved in combining chaotic image encryption and the CNN. These steps include chaotic sequence generation, pixel value mapping, feature extraction, and key management. The algorithm achieves high-strength encryption by performing dissimilarity operations between the chaotic sequence and image pixel values, as well as extracting high-level image features using the CNN. Finally, we conduct experimental evaluations of the algorithm and compare it with traditional chaotic image encryption methods. The experimental results demonstrate that the image encryption algorithm exhibits significant improvements in encryption quality and security while offering advantages in computational performance and encryption/decryption speed.

CapsNetYY1: identifying YY1-mediated chromatin loops based on a capsule network architecture

BackgroundPrevious studies have identified that chromosome structure plays a very important role in gene control. The transcription factor Yin Yang 1 (YY1), a multifunctional DNA binding protein, could form a dimer to mediate chromatin loops and active enhancer-promoter interactions. The deletion of YY1 or point mutations at the YY1 binding sites significantly inhibit the enhancer-promoter interactions and affect gene expression. To date, only a few computational methods are available for identifying YY1-mediated chromatin loops.ResultsWe proposed a novel model named CapsNetYY1, which was based on capsule network architecture to identify whether a pair of YY1 motifs can form a chromatin loop. Firstly, we encode the DNA sequence using one-hot encoding method. Secondly, multi-scale convolution layer is used to extract local features of the sequence, and bidirectional gated recurrent unit is used to learn the features across time steps. Finally, capsule networks (convolution capsule layer and digital capsule layer) used to extract higher level features and recognize YY1-mediated chromatin loops. Compared with DeepYY1, the only prediction for YY1-mediated chromatin loops, our model CapsNetYY1 achieved the better performance on the independent datasets (AUC > 0.99).ConclusionThe results indicate that CapsNetYY1 is an excellent method for identifying YY1-mediated chromatin loops. We believe that the CapsNetYY1 method will be used for predictive classification of other DNA sequences.

Multi-channel EEG emotion recognition through residual graph attention neural network.

In this paper, a novel EEG emotion recognition method based on residual graph attention neural network is proposed. The method constructs a three-dimensional sparse feature matrix according to the relative position of electrode channels, and inputs it into the residual network to extract high-level abstract features containing electrode spatial position information. At the same time, the adjacency matrix representing the connection relationship of electrode channels is constructed, and the time-domain features of multi-channel EEG are modeled using graph. Then, the graph attention neural network is utilized to learn the intrinsic connection relationship between EEG channels located in different brain regions from the adjacency matrix and the constructed graph structure data. Finally, the high-level abstract features extracted from the two networks are fused to judge the emotional state. The experiment is carried out on DEAP data set. The experimental results show that the spatial domain information of electrode channels and the intrinsic connection relationship between different channels contain salient information related to emotional state, and the proposed model can effectively fuse these information to improve the performance of multi-channel EEG emotion recognition.

Text Classification Based on Neural Network Fusion

The goal of text classification is to identify the category to which the text belongs. Text categorization is widely used in email detection, sentiment analysis, topic marking and other fields. However, good text representation is the point to improve the capability of NLP tasks. Traditional text representation adopts bag-of-words model or vector space model, which loses the context information of the text and faces the problems of high latitude and high sparsity,. In recent years, with the increase of data and the improvement of computing performance, the use of deep learning technology to represent and classify texts has attracted great attention. Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and RNN with attention mechanism are used to represent the text, and then to classify the text and other NLP tasks, all of which have better performance than the traditional methods. In this paper, we design two sentence-level models based on the deep network and the details are as follows: (1) Text representation and classification model based on bidirectional RNN and CNN (BRCNN). BRCNN’s input is the word vector corresponding to each word in the sentence; after using RNN to extract word order information in sentences, CNN is used to extract higher-level features of sentences. After convolution, the maximum pool operation is used to obtain sentence vectors. At last, softmax classifier is used for classification. RNN can capture the word order information in sentences, while CNN can extract useful features. Experiments on eight text classification tasks show that BRCNN model can get better text feature representation, and the classification accuracy rate is equal to or higher than that of the prior art. (2) Attention mechanism and CNN (ACNN) model uses the RNN with attention mechanism to obtain the context vector; Then CNN is used to extract more advanced feature information. The maximum pool operation is adopted to obtain a sentence vector; At last, the softmax classifier is used to classify the text. Experiments on eight text classification benchmark data sets show that ACNN improves the stability of model convergence, and can converge to an optimal or local optimal solution better than BRCNN.

A fish image segmentation methodology in aquaculture environment based on multi-feature fusion model

Underwater fish image processing is one of the key technologies to realize intelligent aquaculture. However, due to the complexity of marine environments, underwater fish images usually have the characteristics of color cast, unbalanced contrast, and blur. Current underwater fish image segmentation methods lack adaptive models and have low segmentation accuracy. Hence, this paper constructs a convolutional neural network-based image segmentation model. This paper first proposes a fish image preprocessing method based on pixel threshold segmentation. To enhance the important features in a fish image, this method uses the minimum Euclidean distance between the original image peaks to redefine the threshold and fuses the thresholded image with the original image. Second, to strengthen the extraction of high-level semantic features of images, the multiscale attentional feature extraction module (MAFEM), which fuses the adaptive channel attention mechanism with the hybrid dilated convolutional pyramid pooling module, is proposed. In this paper, a data set in voc format is produced based on underwater fish images, and this data set is used to verify the model in this paper. The mean intersection over union (MIoU) reaches 92.6%. Compared with other traditional segmentation models, the MIoU, mean pixel accuracy (MPA), and balanced F score (F1-score) of the segmentation results of the model in this paper are increased by averages of 1.84%, 0.785%, and 1.18%, respectively. The experimental results show that this model has a better segmentation effect than other models and provides a theoretical basis for intelligent monitoring of underwater fish body length measurement, weight estimation, and discrimination of growth and health statuses.

MGF-net: Multi-channel group fusion enhancing boundary attention for polyp segmentation.

Colonic polyps are the most prevalent neoplastic lesions detected during colorectal cancer screening, and timely detection and excision of these precursor lesions is crucial for preventing multiple malignancies and reducing mortality rates. The pressing need for intelligent polyp detection has led to the development of a high-precision intelligent polyp segmentation network designed to improve polyp screening rates during colonoscopies. In this study, we employed ResNet50 as the backbone network and embedded a multi-channel grouping fusion encoding module in the third to fifth stages to extract high-level semantic features of polyps. Receptive field modules were utilized to capture multi-scale features, and grouping fusion modules were employed to capture salient features in different group channels, guiding the decoder to generate an initial global mapping with improved accuracy. To refine the segmentation of the initial global mapping, we introduced an enhanced boundary weight attention module that adaptively thresholds the initial global mapping using learnable parameters. A self-attention mechanism was then utilized to calculate the long-distance dependency relationship of the polyp boundary area, resulting in an output feature map with enhanced boundaries that effectively refines the boundary of the target area. We carried out contrast experiments of MGF-Net with mainstream polyp segmentation networks on five public datasets of ColonDB, CVC-ColonDB, CVC-612, Kvasir, and ETIS. The results demonstrate that the segmentation accuracy of MGF-Net is significantly improved on the datasets. Furthermore, a hypothesis test was conducted to assess the statistical significance of the computed results. Our proposed MGF-Net outperforms existing mainstream baseline networks and presents a promising solution to the pressing need for intelligent polyp detection. The proposed model is available at https://github.com/xiefanghhh/MGF-NET.

Segmentation of borehole acoustic reflection image using feature pyramid network and transfer learning

Borehole acoustic reflection imaging technology is widely utilized for high-resolution detection of subsurface geological structures. However, the presence of significant noise in migration images poses a challenge for accurate identification of useful reflectors. In this paper, we propose a deep learning-based image segmentation method to address this challenge. Our approach employs a Feature Pyramid Network with an EfficientNet-b4 encoder, which is trained on a large simulation dataset. To handle the class imbalance issue caused by the small proportion of fractures, we employ the Dice loss function. Additionally, the Lovász-hinge loss function is utilized to enhance network training and achieve higher IoU metric values. To extract high-level semantic features, the trained network is fine-tuned on a hand-labeled field dataset using transfer learning. The results obtained from processing field migration images demonstrate the effectiveness and robustness of our proposed methods.

Improving Feature Learning in Remote Sensing Images Using an Integrated Deep Multi-Scale 3D/2D Convolutional Network

Developing complex hyperspectral image (HSI) sensors that capture high-resolution spatial information and voluminous (hundreds) spectral bands of the earth’s surface has made HSI pixel-wise classification a reality. The 3D-CNN has become the preferred HSI pixel-wise classification approach because of its ability to extract discriminative spectral and spatial information while maintaining data integrity. However, HSI datasets are characterized by high nonlinearity, voluminous spectral features, and limited training sample data. Therefore, developing deep HSI classification methods that purely utilize 3D-CNNs in their network structure often results in computationally expensive models prone to overfitting when the model depth increases. In this regard, this paper proposes an integrated deep multi-scale 3D/2D convolutional network block (MiCB) for simultaneous low-level spectral and high-level spatial feature extraction, which can optimally train on limited sample data. The strength of the proposed MiCB model solely lies in the innovative arrangement of convolution layers, giving the network the ability (i) to simultaneously convolve the low-level spectral with high-level spatial features; (ii) to use multiscale kernels to extract abundant contextual information; (iii) to apply residual connections to solve the degradation problem when the model depth increases beyond the threshold; and (iv) to utilize depthwise separable convolutions in its network structure to address the computational cost of the proposed MiCB model. We evaluate the efficacy of our proposed MiCB model using three publicly accessible HSI benchmarking datasets: Salinas Scene (SA), Indian Pines (IP), and the University of Pavia (UP). When trained on small amounts of training sample data, MiCB is better at classifying than the state-of-the-art methods used for comparison. For instance, the MiCB achieves a high overall classification accuracy of 97.35%, 98.29%, and 99.20% when trained on 5% IP, 1% UP, and 1% SA data, respectively.