Average Pooling Layer Research Articles

Automated deep bottleneck residual 82-layered architecture with Bayesian optimization for the classification of brain and common maternal fetal ultrasound planes.

Despite a worldwide decline in maternal mortality over the past two decades, a significant gap persists between low- and high-income countries, with 94% of maternal mortality concentrated in low and middle-income nations. Ultrasound serves as a prevalent diagnostic tool in prenatal care for monitoring fetal growth and development. Nevertheless, acquiring standard fetal ultrasound planes with accurate anatomical structures proves challenging and time-intensive, even for skilled sonographers. Therefore, for determining common maternal fetuses from ultrasound images, an automated computer-aided diagnostic (CAD) system is required. A new residual bottleneck mechanism-based deep learning architecture has been proposed that includes 82 layers deep. The proposed architecture has added three residual blocks, each including two highway paths and one skip connection. In addition, a convolutional layer has been added of size 3 × 3 before each residual block. In the training process, several hyper parameters have been initialized using Bayesian optimization (BO) rather than manual initialization. Deep features are extracted from the average pooling layer and performed the classification. In the classification process, an increase occurred in the computational time; therefore, we proposed an improved search-based moth flame optimization algorithm for optimal feature selection. The data is then classified using neural network classifiers based on the selected features. The experimental phase involved the analysis of ultrasound images, specifically focusing on fetal brain and common maternal fetal images. The proposed method achieved 78.5% and 79.4% accuracy for brain fetal planes and common maternal fetal planes. Comparison with several pre-trained neural nets and state-of-the-art (SOTA) optimization algorithms shows improved accuracy.

DCDS-Net: Deep transfer network based on depth-wise separable convolution with residual connection for diagnosing gastrointestinal diseases

Gastrointestinal (GI) diseases are the most common in the human digestive system and has a significantly higher mortality rate. Accurate evaluation of endoscopic images plays an important role in decision making regarding patient treatment. Recently, convolutional neural networks (CNNs) have been introduced for the diagnosis of GI diseases. However, achieving high accuracy is still a challenging task. To overcome these limitations, we propose the “Densely Connected Depth-wise Separable Convolution-Based Network” (DCDS-Net) model, utilizing depth-wise separable convolution (DWSC) with residual connections and densely connected blocks (DCB), to effectively diagnose various endoscopic images of GI diseases. In addition, we incorporate global average pooling (GAP), batch normalization, dropout and dense layers in DCB to learn rich discriminative features and improve the performance of the model. We explored the feasibility of block-wise fine-tuning using transfer learning on the proposed model to reduce overfitting, and experimentally explore the optimal level of fine-tuning, since transfer learning is well suited to medical data where labeled data is scarce. The proposed method has been evaluated on 6000 labeled endoscopic images containing 4 classes of GI diseases. In addition, data augmentation has been incorporated into the training pipeline to improve the performance of the model. Furthermore, a critical study was conducted to evaluate the generalizability of the proposed model on smaller training samples (e.g., 60 %, 70 %, 80 %, and 90 %). The study employed Grad-CAM to generate heatmaps that identify the regions in the GI tract that are indicative of the presence of different diseases. The results of extensive experiments show that the proposed model shows significant improvements and achieves the highest classification accuracy of 99.33 %, precision of 99.37 %, recall of 99.32 % and outperforms all pre-trained and existing models for the detection of GI diseases. In conclusion, DCDS-Net exhibits high classification performance and can help endoscopists in automatic GI disease diagnosis.

Toward the balance between computational cost and model performance for the void detection of concrete-filled steel tubular structure using one-dimensional Mel-frequency cepstral coefficients and ensemble machine learning

Deep learning (DL) based percussion-acoustic methods have gained attention, but their multi-layer architectures and iterative processes increase computational time and power. This paper proposes a lightweight concrete-filled steel tubular (CFST) void detection method using Mel-frequency cepstral coefficient (MFCC) algorithm and ensemble machine learning. A global averaging pooling (GAP) layer was applied to downscale the two-dimensional MFCC matrix to one-dimensional (1D) time-frequency parameters. Combined with ensemble machine learning, lightweight models that require fewer model parameters, less computation, and smaller memory usage are constructed. Validation experiments were carried out on a CFST specimen with different depths of voids (0, 30, 50, and 80 mm). Robustness tests, considering tap intervals, sampling frequencies, and background noise, were performed by using data augmentation strategies. Two studies investigate feature extraction methods: the selection of the time-frequency analysis algorithm (MFCC or short-time Fourier transform) and the comparison of dimensionality reduction operation (average, maximum, and minimum pooling layers). Comparative analyses with a 1D dilated convolutional neural network (1D dilated CNN) reveal that the MFCC-based ensemble machine learning achieves excellent test accuracy on the raw signal dataset and better generalization capability on the reconstructed signal dataset. Moreover, the ensemble machine learning model exhibits significantly higher computational efficiency compared to DL models. Specifically, the training time of random forest is 17,510 times faster than the 1D dilated CNN. Therefore, the proposed lightweight CFST void detection method, incorporating 1D MFCC parameters and ensemble machine learning, achieves high model accuracy with low computational cost and holds promising potential for practical CFST applications.

EfficientRMT-Net—An Efficient ResNet-50 and Vision Transformers Approach for Classifying Potato Plant Leaf Diseases

The primary objective of this study is to develop an advanced, automated system for the early detection and classification of leaf diseases in potato plants, which are among the most cultivated vegetable crops worldwide. These diseases, notably early and late blight caused by Alternaria solani and Phytophthora infestans, significantly impact the quantity and quality of global potato production. We hypothesize that the integration of Vision Transformer (ViT) and ResNet-50 architectures in a new model, named EfficientRMT-Net, can effectively and accurately identify various potato leaf diseases. This approach aims to overcome the limitations of traditional methods, which are often labor-intensive, time-consuming, and prone to inaccuracies due to the unpredictability of disease presentation. EfficientRMT-Net leverages the CNN model for distinct feature extraction and employs depth-wise convolution (DWC) to reduce computational demands. A stage block structure is also incorporated to improve scalability and sensitive area detection, enhancing transferability across different datasets. The classification tasks are performed using a global average pooling layer and a fully connected layer. The model was trained, validated, and tested on custom datasets specifically curated for potato leaf disease detection. EfficientRMT-Net’s performance was compared with other deep learning and transfer learning techniques to establish its efficacy. Preliminary results show that EfficientRMT-Net achieves an accuracy of 97.65% on a general image dataset and 99.12% on a specialized Potato leaf image dataset, outperforming existing methods. The model demonstrates a high level of proficiency in correctly classifying and identifying potato leaf diseases, even in cases of distorted samples. The EfficientRMT-Net model provides an efficient and accurate solution for classifying potato plant leaf diseases, potentially enabling farmers to enhance crop yield while optimizing resource utilization. This study confirms our hypothesis, showcasing the effectiveness of combining ViT and ResNet-50 architectures in addressing complex agricultural challenges.

A Vision-Based Pose Estimation of a Non-Cooperative Target Based on a Self-Supervised Transformer Network

In the realm of non-cooperative space security and on-orbit service, a significant challenge is accurately determining the pose of abandoned satellites using imaging sensors. Traditional methods for estimating the position of the target encounter problems with stray light interference in space, leading to inaccurate results. Conversely, deep learning techniques require a substantial amount of training data, which is especially difficult to obtain for on-orbit satellites. To address these issues, this paper introduces an innovative binocular pose estimation model based on a Self-supervised Transformer Network (STN) to achieve precise pose estimation for targets even under poor imaging conditions. The proposed method generated simulated training samples considering various imaging conditions. Then, by combining the concepts of convolutional neural networks (CNN) and SIFT features for each sample, the proposed method minimized the disruptive effects of stray light. Furthermore, the feedforward network in the Transformer employed in the proposed method was replaced with a global average pooling layer. This integration of CNN’s bias capabilities compensates for the limitations of the Transformer in scenarios with limited data. Comparative analysis against existing pose estimation methods highlights the superior robustness of the proposed method against variations caused by noisy sample sets. The effectiveness of the algorithm is demonstrated through simulated data, enhancing the current landscape of binocular pose estimation technology for non-cooperative targets in space.

Exploring Symmetry in Digital Image Forensics Using a Lightweight Deep-Learning Hybrid Model for Multiple Smoothing Operators

Digital images are widely used for informal information sharing, but the rise of fake photos spreading misinformation has raised concerns. To address this challenge, image forensics is employed to verify the authenticity and trustworthiness of these images. In this paper, an efficient scheme for detecting commonly used image smoothing operators is presented while maintaining symmetry. A new lightweight deep-learning network is proposed, which is trained with three different optimizers to avoid downsizing to retain critical information. Features are extracted from the activation function of the global average pooling layer in three trained deep networks. These extracted features are then used to train a classification model with an SVM classifier, resulting in significant performance improvements. The proposed scheme is applied to identify averaging, Gaussian, and median filtering with various kernel sizes in small-size images. Experimental analysis is conducted on both uncompressed and JPEG-compressed images, showing superior performance compared to existing methods. Notably, there are substantial improvements in detection accuracy, particularly by 6.50% and 8.20% for 32 × 32 and 64 × 64 images when subjected to JPEG compression at a quality factor of 70.

GraphPhys: Facial video-based physiological measurement with graph neural network

Non-contact facial video-based physiological signal measurement using remote photoplethysmography (rPPG) has shown great potential in many applications (e.g. telemedicine, driver fatigue detection, sleep detection) and has achieved reliable results in the controlled environment. With the development of deep learning, more and more methods have been proposed. However, these methods were usually implemented based on model variations that have achieved better results in computer vision tasks, and finally chose average pooling layer or one-dimensional convolution layer directly reduce dimension. Such these operations ignored the fact that the physiological signals of one individual are consistent in different facial regions. In this paper, we propose a novel remote photoplethysmography framework that utilizes graph neural network to extract physiological signal, named GraphPhys. And we design the Average Relative GraphConv based on the fact that the physiological signals of one individual are always consistent in different facial regions for the remote physiological signal measurement task. We apply GraphPhys to two different architectures for rPPG methods, and the methods based on GraphPhys significantly outperform the original methods on the VIPL-HR (Niu et al., 2019) dataset. To the best of our knowledge, this is the first work to utilize graph neural network to aggregate physiological signals. The code is available at https://github.com/Xiong-JiaHao/GraphPhys.

An adaptive multi-perceptual implicit sampling for hyperspectral and multispectral remote sensing image fusion

Hyperspectral and multispectral remote sensing image fusion (HMIF) can effectively enhance image spatial-spectral information representation. However, existing deep learning algorithms usually use neighboring pixels for interpolation in the sampling process, which ignores the correlation of different receptive field pixels. To solve the above problem, this study proposes an adaptive multi-perceptual field implicit guided sampling generative adversarial network (AMGSGAN) to enhance the long-range perception of the sampling process. The generator consists of precoding, multi-perceptual field feature extraction, and adaptive guided implicit sampling modules. The precoding module encodes the features of the images to enhance the feature expression capability, the multi-perceptual field feature extraction module is used to extract the multi-perceptual field features of the images by setting different convolutional dilation rates, and the adaptive guided implicit sampling module utilizes pixels with different distances to interpolate the features of the pixels to be sampled and fuses the interpolation results adaptively to generate the fusion image. The discriminator incorporates the pure CNN architecture and employs techniques such as batch normalization layer and adaptive average pooling layer to ensure the stability and convergence speed of the network. Comparison experiments on several datasets show that the fusion performance of AMGSGAN is significantly better than the compared algorithms at 4x, 8x and 16x resolutions, which demonstrates the effectiveness of AMGSGAN. The code can be found at https://github.com/chunyuzhu/AMGSGAN.

Abstract 349: Deep Learning Detection of Chest Compression Location Using Transesophageal Echocardiography

Background: Recent work evaluating intra-arrest transesophageal echocardiography (TEE) in cardiac arrest (CA) has demonstrated that chest compression (CC) location is a strong predictor of return of spontaneous circulation. However, advanced skills to interpret images during resuscitation represent an important barrier to implementation of this modality. Deep Learning (DL) models have been increasingly used to perform the automated interpretation of TEE images. Given the unique challenges during the interpretation of images, DL-based automation represents an attractive adjunct to facilitate the use of intra-arrest TEE. We sought to evaluate the feasibility of a DL model to identify CC location in TEE images obtained during CA resuscitation. Methods: We analyzed videos from out-of-hospital and in-hospital CA patients evaluated with TEE collected through the Resuscitative TEE Collaborative Registry (NCT04972526). We reviewed and selected images for the DL algorithm, using midesophageal long axis clips. Videos were selected based on the visualization of anatomical structures necessary to identify the CC location. Images were annotated by an expert classifying them in CC over the LV (CC-LV) vs CC over the left ventricular outflow tract (CC-LVOT). We used TEENet, an end-to-end DL approach with input of TEE clips containing single compression cycles, to classify clips in CC-LV and CC-LVOT. The TEENet architecture includes two 3D convolution layers, followed by a global average pooling layer, and then appends a fully connected layer with a sigmoid layer. We used the F1-score as the evaluation metric for the model's performance. Results: Ten videos containing a total of 47 single CC cycle clips were selected for inclusion in the model, each clip with an average of 25 frames. The model was trained using 80% of the data (37 clips) and tested using the remaining 20% (10 clips). The model achieved an F1-score of 0.7273 in classifying CC-LV from CC-LVOT. Conclusion: We report the creation of a novel DL-model with acceptable precision and recall performance in the detection of CC location using intra-arrest TEE videos. This work is a proof-of-concept for future research aimed to evaluate the potential of DL-assisted TEE-guided CA resuscitation.

An IoT-Based Intelligent Selection of Multidomain Feature for Smart Healthcare Using Reinforcement Learning in Schizophrenia

Currently, the health industry by the Internet of Things (IoT) is developing rapidly, and Electroencephalogram (EEG) signal has become a bridge for human-machine communication. EEG signal feature selection is a key link in brain nerves research and has important practical application value. Therefore, we propose an IoT-based Intelligent Selection of Multi-domain Feature for multimodal transformer EEG signal using reinforcement learning in Schizophrenia. First of all, the features of the multimodal EEG signal sequence by the IoT were extracted from time domain, frequency domain, and spatial domain; Second, the average pooling layer was added to improve the design of the transformer model for deep feature extraction of EEG signal; finally, reinforcement learning was introduced to run operation with the extracted features as input and the improved transformer model as agent. At the same time, entropy and Pearson correlation coefficient calculation were introduced to select feature subsets, and feature intelligent selection of EEG signal in multiple domains was completed through interaction between the agent and the environment. Experiments were performed on DEAP, EEG Motor Movement/Imagery Data set, BCI2008 competition data set and data set of a hospital. The results show that the feature extraction and feature visualization effect of the proposed algorithm is good, and the feature selection precision is as high as about 90%, and the average time is only about 13 seconds. These findings indicate the feasibility of the proposed feature intelligent selection algorithm in the study of EEG signal in schizophrenia, which provides a good basis for further study on the integration of IoT with the healthy industry.

Enhanced capsule network‐based executable files malware detection and classification—deep learning approach

SummaryMalware has considerably increased recently, posing a serious security danger to both people and enterprises. In order to distinguish and stop the negative effects of malware, a variety of machine and deep learning approaches have been used to detect it. However, while extracting malware features, the feature‐to‐feature spatial hierarchy is not taken into account by the existing techniques and as a result, information is lost during the pooling operation. Hence, a modified capsule deep neural network was developed in which discriminative features are extracted from three channel image derived from malware binary with considering feature‐to‐feature spatial hierarchy. Also, conventional capsule deep neural network is modified by adding a global average pooling layer before fully connected layer thereby classified the dataset as malicious or benign without any loss of information. Moreover, these malwares were not accurately classified based on their families using existing variants of convolutional neural network (CNN) since malware family variants can modify due to minute changes in malware binaries. Hence, a hybrid deep convolutional neural network (DCNN) and long‐short‐term memory (LSTM) has been utilized that determine minute changes in malware binaries using LSTM without vanishing gradient issue and effectively perform malware family classification using DCNN. As a result, the proposed approach successfully identifies malware in executable files and categorizes malware into families with 98.5% accuracy.

Investigation of multiple convolutional neural network models on emotion detection

The accurate detection of emotions holds significant importance in the field of psychology, necessitating the careful selection of an appropriate model for facial expression classification. In this study, emotion detection is the classification task to compare the performance of MobileNet, ResNet, and DenseNet. For the detailed model, MobileNet, ResNet50, and DenseNet169 are selected for comparative analysis. The dataset FER-2013 is from Kaggle, which contains a training set and test set consisting of 29, 709 samples and 3589 samples, respectively, with seven facial expression categories. In terms of preprocessing, normalization, and data augmentation are considered. The whole dataset is normalized by dividing 255 and augmented with a Keras image generator. In the model-building step, the structure of the test models is controlled in the same structure. The pre-trained model from the Keras application connects with one global average pooling layer and adds one dense layer at the last as the output layer with the SoftMax activation function. Moreover, this study kept all hyper all parameters the same during the training period. After the model training, the confusion matrix is used to show the class relativity and the loss and accuracy of each model are plotted for analysis. Experimental results demonstrated that the MobileNet achieves 56.08% accuracy on test set which is more competitive than the DenseNet169 and ResNet50 and provides a relatively stable loss.

A deep CNN architecture with novel pooling layer applied to two Sudanese Arabic sentiment data sets

Arabic sentiment analysis has become an important research field in recent years. Initially, work focused on Modern Standard Arabic (MSA), which is the most widely used form. Since then, work has been carried out on several different dialects, including Egyptian, Levantine and Moroccan. Moreover, a number of data sets have been created to support such work. However, up until now, no work has been carried out on Sudanese Arabic, a dialect which has 32 million speakers. In this article, two new public data sets are introduced, the two-class Sudanese Sentiment Data set (SudSenti2) and the three-class Sudanese Sentiment Data set (SudSenti3). In the preparation phase, we establish a Sudanese stopword list. Furthermore, a convolutional neural network (CNN) architecture, Sentiment Convolutional MMA (SCM), is proposed, comprising five CNN layers together with a novel Mean Max Average (MMA) pooling layer, to extract the best features. This SCM model is applied to SudSenti2 and SudSenti3 and shown to be superior to the baseline models, with accuracies of 92.25% and 85.23% (Experiments 1 and 2). The performance of MMA is compared with Max, Avg and Min and shown to be better on SudSenti2, the Saudi Sentiment Data set and the MSA Hotel Arabic Review Data set by 1.00%, 0.83% and 0.74%, respectively (Experiment 3). Next, we conduct an ablation study to determine the contribution to performance of text normalisation and the Sudanese stopword list (Experiment 4). For normalisation, this makes a difference of 0.43% on two-class and 0.45% on three-class. For the custom stoplist, the differences are 0.82% and 0.72%, respectively. Finally, the model is compared with other deep learning classifiers, including transformer-based language models for Arabic, and shown to be comparable for SudSenti2 (Experiment 5).

Dried shiitake mushroom grade recognition using D-VGG network and machine vision.

Grading dried shiitake mushrooms is an indispensable production step, as there are large quality differences between different grades, which affect the product's price and marketability. Dried shiitake mushroom samples have irregular shapes, small morphological differences between different grades of the same species, and they may occur in mixed grades, which causes challenges to the automatic grade recognition using machine vision. In this study, a comprehensive method to solve this problem is provided, including image acquisition, preprocessing, dataset creation, and grade recognition. The osprey optimization algorithm (OOA) is used to improve the computational efficiency of Otsu's threshold binarization and obtain complete mushroom contours samples efficiently. Then, a method for dried shiitake mushroom grade recognition based on the improved VGG network (D-VGG) is proposed. The method uses the VGG16 network as the base framework, optimizes the convolutional layer of the network, and uses a global average pooling layer instead of a fully connected layer to reduce the risk of model overfitting. In addition, a residual module and batch normalization are introduced to enhance the learning effect of texture details, accelerate the convergence of the model, and improve the stability of the training process. An improved channel attention network is proposed to enhance the feature weights of different channels and improve the grading performance of the model. The experimental results show that the improved network model (D-VGG) can recognize different dried shiitake mushroom grades with high accuracy and recognition efficiency, achieving a final grading accuracy of 96.21%, with only 46.77 ms required to process a single image. The dried shiitake mushroom grade recognition method proposed in this study provides a new implementation approach for the dried shiitake mushroom quality grading process, as well as a reference for real-time grade recognition of other agricultural products.

Research on None-Line-of-Sight/Line-of-Sight Identification Method Based on Convolutional Neural Network-Channel Attention Module.

None-Line-of-Sight (NLOS) propagation of Ultra-Wideband (UWB) signals leads to a decrease in the reliability of positioning accuracy. Therefore, it is essential to identify the channel environment prior to localization to preserve the high-accuracy Line-of-Sight (LOS) ranging results and correct or reject the NLOS ranging results with positive bias. Aiming at the problem of the low accuracy and poor generalization ability of NLOS/LOS identification methods based on Channel Impulse Response (CIR) at present, the multilayer Convolutional Neural Networks (CNN) combined with Channel Attention Module (CAM) for NLOS/LOS identification method is proposed. Firstly, the CAM is embedded in the multilayer CNN to extract the time-domain data features of the original CIR. Then, the global average pooling layer is used to replace the fully connected layer for feature integration and classification output. In addition, the public dataset from the European Horizon 2020 Programme project eWINE is used to perform comparative experiments with different structural models and different identification methods. The results show that the proposed CNN-CAM model has a LOS recall of 92.29%, NLOS recall of 87.71%, accuracy of 90.00%, and F1-score of 90.22%. Compared with the current relatively advanced technology, it has better performance advantages.

Novel deep learning models with novel integrated activation functions for autism screening: AutiNet and MinAutiNet

Deep Learning (DL) particularly CNN has the potential to outplay conventional Machine Learning (ML) procedures in the realm of computer vision. With the advent of DL, profound research on autism utilizing fMRI data has declared advancements in autism screening. Yet, DL models are inadequate for heterogeneous datasets encompassing complex fMRI data. To resolve these problems, two cutting-edge AutiNet and MinAutiNet models are proposed with a minimum number of layers and neurons in classifying the autistic subjects. Both AutiNet and MinAutiNet models are the two refined forms of the traditional LeNet-5 model. AutiNet model includes two convolutional layers, two average pooling layers, a flatten layer, and one fully connected layer. One convolutional layer, one average pooling layer, one flatten layer, and one fully connected layer make up the MinAutiNet model. Pre-processing is a pre-requisite phase in the AutiNet and MinAutiNet models where denoising of raw fMRI data is done with slice time correction, realignment, min–max normalization and one hot encoding to obtain numerical values. As inspired by the promising results of non-handcrafted feature extraction, the proposed DL models have been used for feature extraction. Two novel integrated activation functions namely Li-ReLU and S-RReLU are developed and incorporated into the proposed DL models towards improving the performance of autism screening. Both AutiNet and MinAutiNet models surpass the standard accuracy in classifying autistic subjects with minimum number of layers and neurons. Experimental results reveal that the proposed AutiNet attains the maximum accuracy of 77.78% with Caltech and KKI datasets and the proposed MinAutiNet achieves the maximum accuracy of 88.89% with SBL dataset. The proposed models are compared with other state-of-the-art works to ensure its superiority.

Detecting Images in Two-Operator Series Manipulation: A Novel Approach Using Transposed Convolution and Information Fusion

Digital image forensics is a crucial emerging technique, as image editing tools can modify them easily. Most of the latest methods can determine whether a specific operator has edited an image. These methods are suitable for high-resolution uncompressed images. In practice, more than one operator is used to modify image contents repeatedly. In this paper, a reliable scheme using information fusion and deep network networks is presented to recognize manipulation operators and the operator’s series on two operators. A transposed convolutional layer improves the performance of low-resolution JPEG compressed images. In addition, a bottleneck technique is utilized to extend the number of transposed convolutional layers. One average pooling layer is employed to preserve the optimal information flow and evade the overfitting concern among the layers. Moreover, the presented scheme can detect two operator series with various factors without including them in training. The experimental outcomes of the suggested scheme are encouraging and better than the existing schemes due to the availability of sufficient statistical evidence.

State of Charge Estimation of Lithium-Ion Battery Based on IDRSN and BiGRU

Abstract The estimation of state of charge (SOC) is a critical issue in the energy management of electric vehicle (EV) power batteries. However, the current accuracy of SOC estimation methods does not meet the requirements of practical applications. Therefore, this study proposes an improved lithium-ion battery SOC estimation method that combines deep residual shrinkage network (DRSN) and bidirectional gated recurrent unit (BiGRU) to enhance the SOC estimation accuracy. First, we insert the bidirectional gated recurrent unit neural network between the global average pooling layer and the output fully connected layer of the deep residual shrinkage network. This improvement enhances the model’s expressiveness, robustness, and data learning effect. Second, we develop a new activation function called “∂_swish” to replace the original ReLU activation function in the deep residual shrinkage network. The ∂_swish activation function improves the accuracy of the deep network model and reduces the risk of overfitting by utilizing its regularization effect. Finally, we conduct experimental tests at three different temperatures using the FUDS driving cycle dataset and the DST-US06-FUDS continuous driving cycle dataset. The algorithm model’s convergence speed is verified by comparing it with other models. The results show that compared to other models, the proposed method significantly improves SOC estimation accuracy at three different temperatures. In addition, the method demonstrates a high convergence speed.

Fine-tuning U-net for medical image segmentation based on activation function, optimizer and pooling layer

<span lang="EN-US">U-net convolutional neural network (CNN) is a famous architecture developed to deal with medical images. Fine-tuning CNNs is a common technique used to enhance their performance by selecting the building blocks which can provide the ultimate results. This paper introduces a method for tuning U-net architecture to improve its performance in medical image segmentation. The experiment is conducted using an x-ray image segmentation approach. The performance of U-net CNN in lung x-ray image segmentation is studied with different activation functions, optimizers, and pooling-bottleneck-layers. The analysis focuses on creating a method that can be applied for tuning U-net, like CNNs. It also provides the best activation function, optimizer, and pooling layer to enhance U-net CNN’s performance on x-ray image segmentation. The findings of this research showed that a U-net architecture worked supremely when we used the LeakyReLU activation function and average pooling layer as well as RMSProb optimizer. The U-net model accuracy is raised from 89.59 to 93.81% when trained and tested with lung x-ray images and uses the LeakyReLU activation function, average pooling layer, and RMSProb optimizer. The fine-tuned model also enhanced accuracy results with three other datasets.</span>

Natural gas pipeline leak detection based on acoustic signal analysis and feature reconstruction

The natural gas pipeline leakage detection task based on acoustic signal has some problems such as background noise coverage, lack of effective features, and low fault identification accuracy caused by small sample data. However, only one of these problems was usually studied in previous technologies. Almost no one has attempted to challenge multiple issues at the same time. In this study, a natural gas pipeline leak detection model that integrates acoustic feature processing techniques and feature reconstruction is proposed to resolve the above problems collaboratively. This model consists of two components. The first component is a feature processing technique of the acoustic signal that integrates frequency domain vector denoising and time domain associative function feature enhancement. The second component is a one-dimensional convolutional neural network with an expanded structural feature encoder (FAE) for feature reconstruction (FAE-1D-CNN). In the feature processing stage of the acoustic signal, firstly, the acoustic signal collected by the acoustic sensor is discretized into a digital signal. Secondly, the energy modal function is used to perform high/low energy modal clustering of digital signal features. The feature validity is enhanced by adding association factors to the low-energy modal features matrix. A low-pass filtering method is then used in the high-energy modal features to remove the background noise coverage of the high-frequency components. In the fault feature extraction stage, a feature encoder (FAE) is introduced in the 1D-CNN network to extract effective fault features while performing secondary reconstruction of local spatial features, addressing the problem of small sample leakage signals with few effective fault characteristics. The global average pooling layer is used instead of the fully connected layer, and the Softmax function is adopted as the classifier for fault discrimination. The performance of the proposed method was evaluated on the GPLA-12 dataset, and the fault identification accuracy is up to 95.17%. Compared with other competing methods, the method in this paper exhibits optimal performance and has broad application prospects.