Traditional Convolutional Neural Network Research Articles

Learning Lightweight Dynamic Kernels With Attention Inside via Local-Global Context Fusion.

Traditional convolutional neural networks (CNNs) share their kernels among all positions of the input, which may constrain the representation ability in feature extraction. Dynamic convolution proposes to generate different kernels for different inputs to improve the model capacity. However, the total parameters of the dynamic network can be significantly huge. In this article, we propose a lightweight dynamic convolution method to strengthen traditional CNNs with an affordable increase of total parameters and multiply-adds. Instead of generating the whole kernels directly or combining several static kernels, we choose to "look inside," learning the attention within convolutional kernels. An extra network is used to adjust the weights of kernels for every feature aggregation operation. By combining local and global contexts, the proposed approach can capture the variance among different samples, the variance in different positions of the feature maps, and the variance in different positions inside sliding windows. With a minor increase in the number of model parameters, remarkable improvements in image classification on CIFAR and ImageNet with multiple backbones have been obtained. Experiments on object detection also verify the effectiveness of the proposed method.

Enhancing Video Anomaly Detection Using a Transformer Spatiotemporal Attention Unsupervised Framework for Large Datasets

This work introduces an unsupervised framework for video anomaly detection, leveraging a hybrid deep learning model that combines a vision transformer (ViT) with a convolutional spatiotemporal relationship (STR) attention block. The proposed model addresses the challenges of anomaly detection in video surveillance by capturing both local and global relationships within video frames, a task that traditional convolutional neural networks (CNNs) often struggle with due to their localized field of view. We have utilized a pre-trained ViT as an encoder for feature extraction, which is then processed by the STR attention block to enhance the detection of spatiotemporal relationships among objects in videos. The novelty of this work is utilizing the ViT with the STR attention to detect video anomalies effectively in large and heterogeneous datasets, an important thing given the diverse environments and scenarios encountered in real-world surveillance. The framework was evaluated on three benchmark datasets, i.e., the UCSD-Ped2, CHUCK Avenue, and ShanghaiTech. This demonstrates the model’s superior performance in detecting anomalies compared to state-of-the-art methods, showcasing its potential to significantly enhance automated video surveillance systems by achieving area under the receiver operating characteristic curve (AUC ROC) values of 95.6, 86.8, and 82.1. To show the effectiveness of the proposed framework in detecting anomalies in extra-large datasets, we trained the model on a subset of the huge contemporary CHAD dataset that contains over 1 million frames, achieving AUC ROC values of 71.8 and 64.2 for CHAD-Cam 1 and CHAD-Cam 2, respectively, which outperforms the state-of-the-art techniques.

SMARTSeiz: Deep Learning With Attention Mechanism for Accurate Seizure Recognition in IoT Healthcare Devices.

The Internet of Things (IoT) is capable of controlling the healthcare monitoring system for remote-based patients. Epilepsy, a chronic brain syndrome characterized by recurrent, unpredictable attacks, affects individuals of all ages. IoT-based seizure monitoring can greatly enhance seizure patients' quality of life. IoT device acquires patient data and transmits it to a computer program so that doctors can examine it. Currently, doctors invest significant manual effort in inspecting Electroencephalograph (EEG) signals to identify seizure activity. However, EEG-based seizure detection algorithms face challenges in real-world scenarios due to non-stationary EEG data and variable seizure patterns among patients and recording sessions. Therefore, a sophisticated computer-based approach is necessary to analyze complex EEG records. In this work, the authors proposed a hybrid approach by combining traditional convolution neural (CN) and recurrent neural networks (RNN) along with an attention mechanism for the automatic recognition of epileptic seizures through EEG signal analysis. This attention mechanism focuses on significant subsets of EEG data for class recognition, resulting in improved model performance. The proposed methods are evaluated using a publicly available UCI epileptic seizure recognition dataset, which consists of five classes: four normal conditions and one abnormal seizure condition. Experimental results demonstrate that the suggested approach achieves an overall accuracy of 97.05% for the five-class EEG recognition data, with an accuracy of 99.52% for binary classification distinguishing seizure cases from normal instances. Furthermore, the proposed intelligent seizure recognition model is compatible with an IoMT (Internet of Medical Things) cloud-based smart healthcare framework.

TRAFFIC SIGN RECOGNITION IN CHALLENGING WEATHER CONDITIONS USING CONVOLUTIONAL NEURAL NETWORKS

The ability of autonomous driving systems to recognize traffic signs in a variety of environmental conditions is crucial to their reliability. This study uses convolutional neural networks (CNNs) to provide a novel method for improving the accuracy of traffic sign recognition systems under challenging weather situations. The research focuses on developing a CNN-based model that is trained using the augmented German Traffic Sign Recognition Benchmark (GTSRB) dataset, which contains over 50,000 labeled images of road signs across 43 categories. To attempt to overcome the limitations of traditional CNN models under adverse environmental conditions, this work presents adaptive feature extraction layers that are especially made to reduce visibility problems brought on by rain, fog, and snow. By taking a comprehensive approach, the model uses advanced data augmentation methods to simulate different weather scenarios, greatly increasing the diversity of the training dataset. Through an analysis of theoretical and practical aspects, the study demonstrates how CNNs enhance the accuracy and efficiency of road sign detection systems in a different weather condition. Research evaluates the model's effectiveness using metrics such as precision, recall, and F1-score, proving its capability to reduce false positives and accurately detect relevant road sign instances. Additionally, the paper highlights the importance of careful dataset preparation and augmentation, model optimization, and training enhancements to improve the performance of road sign detection systems. The results of this research offer benefits for intelligent transport systems, autonomous driving, and road safety, pointing towards further advancements in precise and dependable road sign recognition technology.

Comparative Study of Adversarial Defenses: Adversarial Training and Regularization in Vision Transformers and CNNs

Transformer-based models are driving a significant revolution in the field of machine learning at the moment. Among these innovations, vision transformers (ViTs) stand out for their application of transformer architectures to vision-related tasks. By demonstrating performance as good, if not better, than traditional convolutional neural networks (CNNs), ViTs have managed to capture considerable interest in the field. This study focuses on the resilience of ViTs and CNNs in the face of adversarial attacks. Such attacks, which introduce noise into the input of machine learning models to produce incorrect outputs, pose significant challenges to the reliability of machine learning models. Our analysis evaluated the adversarial robustness of CNNs and ViTs by using regularization techniques and adversarial training methods. Adversarial training, in particular, represents a traditional approach to boosting defenses against these attacks. Despite its prominent use, our findings reveal that regularization techniques enable vision transformers and, in most cases, CNNs to enhance adversarial defenses more effectively. Through testing datasets like CIFAR-10 and CIFAR-100, we demonstrate that vision transformers, especially when combined with effective regularization strategies, demonstrate adversarial robustness, even without adversarial training. Two main inferences can be drawn from our findings. Firstly, it emphasizes how effectively vision transformers could strengthen artificial intelligence defenses against adversarial attacks. Secondly, it shows how regularization, which requires much fewer computational resources and covers a wide range of adversarial attacks, can be effective for adversarial defenses. Understanding and improving a model’s resilience to adversarial attacks is crucial for developing secure, dependable systems that can handle the complexity of real-world applications as artificial intelligence and machine learning technologies advance.

Clivia biosensor: Soil moisture identification based on electrophysiology signals with deep learning

Research has shown that plants have the ability to detect environmental changes and generate electrical signals in response. These electrical signals can regulate the physiological state of plants and produce corresponding feedback. This suggests that plants have the potential to be used as biosensors for monitoring environmental information. However, there are current challenges in linking environmental information with plant electrical signals, especially in collecting and classifying the corresponding electrical signals under soil moisture gradients. This study documented the electrical signals of clivia under different soil moisture gradients and created a dataset for classifying electrical signals. Subsequently, we proposed a lightweight convolutional neural network (CNN) model (PlantNet) for classifying the electrical signal dataset. Compared to traditional CNN models, our model achieved optimal classification performance with the lowest computational resource consumption. The model achieved an accuracy of 99.26%, precision of 99.31%, recall of 92.26%, F1-score of 99.21%, with 0.17M parameters, a size of 7.17MB, and 14.66M FLOPs. Therefore, this research provides scientific evidence for the future development of plants as biosensors for detecting soil moisture, and offers insight into developing plants as biosensors for detecting signals such as ozone, PM2.5, Volatile Organic Compounds(VOCs), and more. These studies are expected to drive the development of environmental monitoring technology and provide new pathways for better understanding the interaction between plants and the environment.

Counterfeit Currency Detection Using Machine Learning

Counterfeit currency detection is a critical aspect of maintaining financial integrity. This paper introduces an innovative methodology for identifying counterfeit Indian rupee notes. By employing advanced image processing techniques and Convolutional Neural Networks (CNNs), specifically ResNet architecture, the proposed system achieves impressive accuracy in distinguishing genuine from counterfeit banknotes. The dataset, meticulously compiled through a combination of web scraping and synthetic printing, ensures robust training and evaluation. Experimental findings underscore the efficacy of the ResNet-based approach, highlighting its superiority over traditional CNN models. This research contributes significantly to bolstering currency security and trust in financial transactions, offering practical insights for combating counterfeit currency challenges.

Analyzing variation of water inflow to inland lakes under climate change: Integrating deep learning and time series data mining

The alarming depletion of global inland lakes in recent decades makes it essential to predict water inflow from rivers to lakes (WIRL) trend and unveil the dominant influencing driver, particularly in the context of climate change. The raw time series data contains multiple components (i.e., long-term trend, seasonal periodicity, and random noise), which makes it challenging for traditional machine/deep learning techniques to effectively capture long-term trend information. In this study, a novel FactorConvSTLnet (FCS) method is developed through integrating STL decomposition, convolutional neural networks (CNN), and factorial analysis into a general framework. FCS is more robust in long-term WIRL trend prediction through separating trend information as a modeling predictor, as well as unveiling predominant drivers. FCS is applied to typical inland lakes (the Aral Sea and the Lake Balkhash) in Central Asia, and results indicate that FCS (Nash-Sutcliffe efficiency = 0.88, root mean squared error = 67m³/s, mean relative error = 10%) outperforms the traditional CNN. Some main findings are: (i) during 1960–1990, reservoir water storage (WSR) was the dominant driver for the two lakes, respectively contributing to 71% and 49%; during 1991–2014 and 2015–2099, evaporation (EVAP) would be the dominant driver, with the contribution of 30% and 47%; (ii) climate change would shift the dominant driver from human activities to natural factors, where EVAP and surface snow amount (SNW) have an increasing influence on WIRL; (iii) compared to SSP1-2.6, the SNW contribution would decrease by 26% under SSP5-8.5, while the EVAP contribution would increase by 9%. The findings reveal the main drivers of shrinkage of the inland lakes and provide the scientific basis for promoting regional ecological sustainability.

Video-Based Sign Language Recognition via ResNet and LSTM Network.

Sign language recognition technology can help people with hearing impairments to communicate with non-hearing-impaired people. At present, with the rapid development of society, deep learning also provides certain technical support for sign language recognition work. In sign language recognition tasks, traditional convolutional neural networks used to extract spatio-temporal features from sign language videos suffer from insufficient feature extraction, resulting in low recognition rates. Nevertheless, a large number of video-based sign language datasets require a significant amount of computing resources for training while ensuring the generalization of the network, which poses a challenge for recognition. In this paper, we present a video-based sign language recognition method based on Residual Network (ResNet) and Long Short-Term Memory (LSTM). As the number of network layers increases, the ResNet network can effectively solve the granularity explosion problem and obtain better time series features. We use the ResNet convolutional network as the backbone model. LSTM utilizes the concept of gates to control unit states and update the output feature values of sequences. ResNet extracts the sign language features. Then, the learned feature space is used as the input of the LSTM network to obtain long sequence features. It can effectively extract the spatio-temporal features in sign language videos and improve the recognition rate of sign language actions. An extensive experimental evaluation demonstrates the effectiveness and superior performance of the proposed method, with an accuracy of 85.26%, F1-score of 84.98%, and precision of 87.77% on Argentine Sign Language (LSA64).

Video Anomaly Detection Utilizing Efficient Spatiotemporal Feature Fusion with 3D Convolutions and Long Short‐Term Memory Modules

Surveillance cameras produce vast amounts of video data, posing a challenge for analysts due to the infrequent occurrence of unusual events. To address this, intelligent surveillance systems leverage AI and computer vision to automatically detect anomalies. This study proposes an innovative method combining 3D convolutions and long short‐term memory (LSTM) modules to capture spatiotemporal features in video data. Notably, a structured coarse‐level feature fusion mechanism enhances generalization and mitigates the issue of vanishing gradients. Unlike traditional convolutional neural networks, the approach employs depth‐wise feature stacking, reducing computational complexity and enhancing the architecture. Additionally, it integrates microautoencoder blocks for downsampling, eliminates the computational load of ConvLSTM2D layers, and employs frequent feature concatenation blocks during upsampling to preserve temporal information. Integrating a Conv‐LSTM module at the down‐ and upsampling stages enhances the model's ability to capture short‐ and long‐term temporal features, resulting in a 42‐layer network while maintaining robust performance. Experimental results demonstrate significant reductions in false alarms and improved accuracy compared to contemporary methods, with enhancements of 2.7%, 0.6%, and 3.4% on the UCSDPed1, UCSDPed2, and Avenue datasets, respectively.

A Study on the Accuracy of Micro Expression Based Deception Detection with Hybrid Deep Neural Network Models

This article details a study on enhancing deception detection accuracy by using Hybrid Deep Neural Network (HDNN) models. The research, focusing on fear-related micro-expressions, utilizes a diverse dataset of responses to high-stakes questions. It analyzes facial action units (AUs) and pupil size variations through data preprocessing and feature extraction. The HDNN model outperforms the traditional Convolutional Neural Network (CNN) with a 91% accuracy rate. The findings’ implications for security, law enforcement, psychology, and behavioral treatments are discussed. Ethical considerations of deception detection technology deployment and future research directions, including cross-cultural studies, real-world assessments, ethical guidelines, studies on emotional expression dynamics, “explainable AI” development, and multimodal data integration, are also explored. The study contributes to deception detection knowledge and highlights the potential of machine learning techniques, especially HDNN, in improving decision-making and security in high-stakes situations.

A novel fish individual recognition method for precision farming based on knowledge distillation strategy and the range of the receptive field.

With the continuous development of green and high-quality aquaculture technology, the process of industrialized aquaculture has been promoted. Automation, intelligence, and precision have become the future development trend of the aquaculture industry. Fish individual recognition can further distinguish fish individuals based on the determination of fish categories, providing basic support for fish disease analysis, bait feeding, and precision aquaculture. However, the high similarity of fish individuals and the complexity of the underwater environment presents great challenges to fish individual recognition. To address these problems, we propose a novel fish individual recognition method for precision farming that rethinks the knowledge distillation strategy and the chunking method in the vision transformer. The method uses the traditional convolutional neural network model as the teacher model, introducing the teacher token to guide the student model to learn the fish texture features. We propose stride patch embedding to expand the range of the receptive field, thus enhancing the local continuity of the image, and self-attention-pruning to discard unimportant tokens and reduce the model computation. The experimental results on the DlouFish dataset show that the proposed method in this paper improves accuracy by 3.25% compared to ECA Resnet152, with an accuracy of 93.19%, and also outperforms other vision transformer models.

Classification of Muscular Dystrophies from MR Images Improves Using the Swin Transformer Deep Learning Model.

Muscular dystrophies present diagnostic challenges, requiring accurate classification for effective diagnosis and treatment. This study investigates the efficacy of deep learning methodologies in classifying these disorders using skeletal muscle MRI scans. Specifically, we assess the performance of the Swin Transformer (SwinT) architecture against traditional convolutional neural networks (CNNs) in distinguishing between healthy individuals, Becker muscular dystrophy (BMD), and limb-girdle muscular Dystrophy type 2 (LGMD2) patients. Moreover, 3T MRI scans from a retrospective dataset of 75 scans (from 54 subjects) were utilized, with multiparametric protocols capturing various MRI contrasts, including T1-weighted and Dixon sequences. The dataset included 17 scans from healthy volunteers, 27 from BMD patients, and 31 from LGMD2 patients. SwinT and CNNs were trained and validated using a subset of the dataset, with the performance evaluated based on accuracy and F-score. Results indicate the superior accuracy of SwinT (0.96), particularly when employing fat fraction (FF) images as input; it served as a valuable parameter for enhancing classification accuracy. Despite limitations, including a modest cohort size, this study provides valuable insights into the application of AI-driven approaches for precise neuromuscular disorder classification, with potential implications for improving patient care.

A bearing fault diagnosis approach based on an improved neural network combined with transfer learning

Abstract In modern industrial systems, bearing failures account for 30%–40% of industrial machinery faults. Traditional convolutional neural network suffers from gradient vanishing and overfitting, resulting in a poor diagnostic accuracy. To address the issues, a new bearing fault diagnosis approach was proposed based on an improved AlexNet neural network combined with transfer learning. After decomposition and noise-reduction, reconstructed vibration signals were transformed into 2D images, then input into the improved AlexNet for training and follow-up transfer learning. Program auto-tuning and image-enhancing techniques were employed to increase the diagnostic accuracy in this study. The approach was verified with the datasets from Case Western Reserve University (CWRU), Jiangnan University (JNU), and the Association for Mechanical Failure Prevention Technology (MFPT). The results showed that the diagnostic accuracies by normal learning were more than 97% for CWRU and JNU datasets, and 100% for MFPT dataset. After transfer learning, the accuracies all reached above 99.5%. The proposed approach was demonstrated to be able to effectively diagnose the bearing faults.

DualTrans: A Novel Glioma Segmentation Framework Based on a Dual-Path Encoder Network and Multi-View Dynamic Fusion Model

Segmentation methods based on convolutional neural networks (CNN) have achieved remarkable results in the field of medical image segmentation due to their powerful representation capabilities. However, for brain-tumor segmentation, owing to the significant variations in shape, texture, and location, traditional convolutional neural networks (CNNs) with limited convolutional kernel-receptive fields struggle to model explicit long-range (global) dependencies, thereby restricting segmentation accuracy and making it difficult to accurately identify tumor boundaries in medical imaging. As a result, researchers have introduced the Swin Transformer, which has the capability to model long-distance dependencies, into the field of brain-tumor segmentation, offering unique advantages in the global modeling and semantic interaction of remote information. However, due to the high computational complexity of the Swin Transformer and its reliance on large-scale pretraining, it faces constraints when processing large-scale medical images. Therefore, this study addresses this issue by proposing a smaller network, consisting of a dual-encoder network, which also resolves the instability issue that arises in the training process of large-scale visual models with the Swin Transformer, where activation values of residual units accumulate layer by layer, leading to a significant increase in differences in activation amplitudes across layers and causing model instability. The results of the experimental validation using real data show that our dual-encoder network has achieved significant performance improvements, and it also demonstrates a strong appeal in reducing computational complexity.

ST-MDAMNet: Swin transformer combines multi-dimensional attention mechanism for semantic segmentation of high-resolution earth surface images

In the derection of remote sensing (RS) image analysis, semantic segmentation, as an important technology, is of key significance for the identification and analysis of land surface cover types. In recent years, applying deep learning models to tasks such as road extraction, water distribution extraction, building classification and building segmentation from RS images has become an important research hotspot. Due to its limited receptive field, traditional convolutional neural networks (CNN) cannot effectively capture global context information. Transformer uses the multi-head self-attention mechanism to capture a wide range of information and can solve this problem well. Therefore, we proposed ST-MDAMNet based on Swin Transformer and combined with the multi-dimensional attention mechanism. First, a feature enhancement module (FAM) is introduced after each stage of the Swin Transformer encoder to effectively enhance the model’s proficiency in identifying essential information. Secondly, a feature fusion module (FFM) is proposed to effectively fuse the multi-scale information of the encoder part. It further improves the expression ability of different dimensional features and effectively improves the detection effect of small targets. Ultimately, the fused features are input into the multi-dimensional attention module (MDAM) to carefully optimize the features, which greatly increases the effect of semantic segmentation of RS images. We demonstrate the effectiveness of each module through ablation experiments. Comparative experiments are completed on two publicly large-scale datasets, and the proposed method shows excellent results compared with state-of-the-art methods.

A lightweight convolutional neural network for working condition intelligent diagnosis of pumping units

Accurate diagnosis of pumping unit working conditions and prompt fault resolution are vital for ensuring safe production in oilfields. In response to the limitations of traditional convolutional neural networks in diagnosing pumping unit working conditions, a lightweight convolutional neural network is proposed. This network addresses the challenges of shallow networks, which cannot effectively extract indicator diagram features, and deep networks, which suffer from excessive parameterization and high redundancy. Firstly, the combination of depthwise separable convolution and dilated convolution efficiently extracts the contour features of the indicator diagram while reducing the model’s parameter count. Additionally, the Attention Embedded Atrous Spatial Pyramid Pooling and the coordinate attention mechanism are incorporated to further capture the shape and position information of indicator diagram lines. The experimental results demonstrate that the proposed lightweight network has 1.26 M parameters and storage size of 4.82 M. Compared to ResNet-50, VGG-16, and MobileNet-V2, its parameter count and storage size are approximately 1/18, 1/11, and 1/2, respectively, making it easily deployable on hardware-limited diagnostic platforms. Moreover, with an average diagnostic accuracy of 98.17%, surpassing existing networks, it enables more effective diagnosis of pumping unit working conditions, thereby enhancing the reliability and accuracy of pumping unit operational monitoring.

Graph Kernel Neural Networks.

The convolution operator at the core of many modern neural architectures can effectively be seen as performing a dot product between an input matrix and a filter. While this is readily applicable to data such as images, which can be represented as regular grids in the Euclidean space, extending the convolution operator to work on graphs proves more challenging, due to their irregular structure. In this article, we propose to use graph kernels, i.e., kernel functions that compute an inner product on graphs, to extend the standard convolution operator to the graph domain. This allows us to define an entirely structural model that does not require computing the embedding of the input graph. Our architecture allows to plug-in any type of graph kernels and has the added benefit of providing some interpretability in terms of the structural masks that are learned during the training process, similar to what happens for convolutional masks in traditional convolutional neural networks (CNNs). We perform an extensive ablation study to investigate the model hyperparameters' impact and show that our model achieves competitive performance on standard graph classification and regression datasets.

Multiscale knowledge distillation with attention based fusion for robust human activity recognition

Knowledge distillation is an effective approach for training robust multi-modal machine learning models when synchronous multimodal data are unavailable. However, traditional knowledge distillation techniques have limitations in comprehensively transferring knowledge across modalities and models. This paper proposes a multiscale knowledge distillation framework to address these limitations. Specifically, we introduce a multiscale semantic graph mapping (SGM) loss function to enable more comprehensive knowledge transfer between teacher and student networks at multiple feature scales. We also design a fusion and tuning (FT) module to fully utilize correlations within and between different data types of the same modality when training teacher networks. Furthermore, we adopt transformer-based backbones to improve feature learning compared to traditional convolutional neural networks. We apply the proposed techniques to multimodal human activity recognition and compared with the baseline method, it improved by 2.31% and 0.29% on the MMAct and UTD-MHAD datasets. Ablation studies validate the necessity of each component.

Neural Circuit Policies for Virtual Character Control

The development of high-stakes decision-making neural agents that interact with complex environments, such as video games, is an important aspect of AI research with numerous potential applications. Reinforcement learning combined with deep learning architectures (DRL) has shown remarkable success in various genres of games. The performance of DRL is heavily dependent upon the neural networks resides within them. Although these algorithms perform well in offline testing but the performance deteriorates in noisy and sub-optimal conditions, creating safety and security issues. To address these, we propose a hybrid deep learning architecture that combines a traditional convolutional neural network with worm brain-inspired neural circuit policies. This allows the agent to learn key coherent features from the environment and interpret its dynamics. The obtained DRL agent was not only able to achieve an optimal policy quickly, but it was also the most noise-resilient with the highest success rate. Our research indicates that only 20 control neurons (12 inter-neurons and 8 command neurons) are sufficient to achieve competitive results. We implemented and analyzed the agent in the popular video game Doom, demonstrating its effectiveness in practical applications.