LeNet-5 Network Research Articles

A lightweight network architecture for traffic sign recognition based on enhanced LeNet-5 network.

As an important part of the unmanned driving system, the detection and recognition of traffic sign need to have the characteristics of excellent recognition accuracy, fast execution speed and easy deployment. Researchers have applied the techniques of machine learning, deep learning and image processing to traffic sign recognition successfully. Considering the hardware conditions of the terminal equipment in the unmanned driving system, in this research work, the goal was to achieve a convolutional neural network (CNN) architecture that is lightweight and easily implemented for an embedded application and with excellent recognition accuracy and execution speed. As a classical CNN architecture, LeNet-5 network model was chosen to be improved, including image preprocessing, improving spatial pool convolutional neural network, optimizing neurons, optimizing activation function, etc. The test experiment of the improved network architecture was carried out on German Traffic Sign Recognition Benchmark (GTSRB) database. The experimental results show that the improved network architecture can obtain higher recognition accuracy in a short interference time, and the algorithm loss is significantly reduced with the progress of training. At the same time, compared with other lightweight network models, this network architecture gives a good recognition result, with a recognition accuracy of 97.53%. The network structure is simple, the algorithm complexity is low, and it is suitable for all kinds of terminal equipment, which can have a wider application in unmanned driving system.

Application of deep learning classification model for regional evaluation of roof pressure support evolution effects over time in coal mining face

Hydraulic support leg pressure serves as a crucial indicator for assessing work face support quality. Current evaluation methods for support quality primarily concentrate on static analyses-like inadequate initial support force, pressure overrun, and uneven bracket force-while neglecting dynamic column pressure changes. This paper introduces a model for assessing hydraulic support quality using deep learning techniques. Real-time data is preprocessed into a spatio-temporal pressure sub-matrix sample, which is then inputted into the model. This process assesses the support quality type and characterizes its dynamic evolution within the area. The model facilitates the identification of dynamic support quality effects in the working face area, aiding operators in making targeted adjustments to hydraulic support status. Experimental results revealed that the optimized LeNet-5 network-adjusting parameters like convolutional layer count, kernel size, and ReLU activation function-achieved the highest classification accuracy of 85.25 % for support quality, surpassed other networks. Furthermore, the improved LeNet-5 network outperformed other networks in both F1 score and recall. Additionally, the improved LeNet-5 network achieved faster convergence to the optimal solution, accelerated training speed. This highlighted its advantages in evaluating the spatio-temporal support quality of hydraulic supports in smart mining operations.

Railway wagon bearing fault diagnosis method based on improved sparrow search algorithm optimizing variational mode decomposition and multi-level convolutional neural network.

Ensuring the safe operation of trains hinges on precise bearing condition monitoring, given the pivotal role bearings play in railway wagons. The status and maintenance of wagon bearings are of paramount concern, necessitating a shift from traditional maintenance approaches reliant on schedules and experience, which often lack real-time precision and efficiency. To address this challenge, our research focuses on enhancing the sparrow search algorithm by incorporating logistic chaos mapping and the levy flight strategy. This enhanced algorithm optimizes variational mode decomposition parameters, utilizing intrinsic mode components' average dispersion entropy as the fitness function. This optimization is integrated with a multi-level convolutional neural network for bearing fault diagnosis. Our findings demonstrate the improved algorithm's enhanced spatial search capabilities and reduced modal aliasing in the frequency components. Experimental validation on public datasets and the group's experimental platform for railway wagons shows that multi-level convolutional neural networks have higher diagnostic accuracy and faster convergence speeds than traditional models such as LeNet-5, AlexNet, and convolutional neural network. Our research introduces a highly accurate and widely applicable methodology for mechanical equipment fault diagnosis, aligning with the requirements of the "smart" era.

Enabling Efficient On-Edge Spiking Neural Network Acceleration with Highly Flexible FPGA Architectures

Spiking neural networks (SNNs) promise to perform tasks currently performed by classical artificial neural networks (ANNs) faster, in smaller footprints, and using less energy. Neuromorphic processors are set out to revolutionize computing at a large scale, but the move to edge-computing applications calls for finely-tuned custom implementations to keep pushing towards more efficient systems. To that end, we examined the architectural design space for executing spiking neuron models on FPGA platforms, focusing on achieving ultra-low area and power consumption. This work presents an efficient clock-driven spiking neuron architecture used for the implementation of both fully-connected cores and 2D convolutional cores, which rely on deep pipelines for synaptic processing and distributed memory for weight and neuron states. With them, we developed an accelerator for an SNN version of the LeNet-5 network trained on the MNIST dataset. At around 5.5 slices/neuron and only 348 mW, it is able to use 33% less area and four times less power per neuron as current state-of-the-art implementations while keeping low simulation step times.

Assessing Efficiency in Artificial Neural Networks

The purpose of this work was to develop an assessment technique and subsequent metrics that help in developing an understanding of the balance between network size and task performance in simple model networks. Here, exhaustive tests on simple model neural networks and datasets are used to validate both the assessment approach and the metrics derived from it. The concept of neural layer state space is introduced as a simple mechanism for understanding layer utilization, where a state is the on/off activation state of all neurons in a layer for an input. Neural efficiency is computed from state space to measure neural layer utilization, and a second metric called the artificial intelligence quotient (aIQ) was created to balance neural network performance and neural efficiency. To study aIQ and neural efficiency, two simple neural networks were trained on MNIST: a fully connected network (LeNet-300-100) and a convolutional neural network (LeNet-5). The LeNet-5 network with the highest aIQ was 2.32% less accurate but contained 30,912 times fewer parameters than the network with the highest accuracy. Both batch normalization and dropout layers were found to increase neural efficiency. Finally, networks with a high aIQ are shown to be resistant to memorization and overtraining as well as capable of learning proper digit classification with an accuracy of 92.51%, even when 75% of the class labels are randomized. These results demonstrate the utility of aIQ and neural efficiency as metrics for determining the performance and size of a small network using exemplar data.

Hydrogen leakage diagnosis of fuel cell vehicles based on a dynamic-static recognition approach with dual-channel LeNet network

The diagnosis of hydrogen leakage has attracted increasing attention, especially in fuel cell vehicles with high-pressure hydrogen systems within which the chance of deflagration is pretty high. Although online intelligent diagnosis can be achieved with the pressure signal of high-pressure hydrogen tank, the feature extraction process of machine learning is a time-consuming and laborious work and greatly impacts the final result. Motivated by this fact, a dynamic-static recognition approach and a dual-channel LeNet network based on transfer learning and D-S evidence theory are proposed to diagnose the hydrogen leakage of vehicles. The proposed method has the advantages of automatic feature extraction, high accuracy, and real-time recognition capability. The major contributions lie in two aspects. On the one hand, through the dynamic-static recognition approach, pressure signals of high-pressure hydrogen tank are converted into graphic characteristics: Markov transition field (MTF) and Gramian angular field (GAF). On the other hand, for the proposed dual-channel LeNet network, it uses transfer learning to reduce the training loss and improve the accuracy of the LeNet channel, while the D-S evidence theory is used to amalgamate the dual-channel LeNet outputs to obtain highly accurate hydrogen leak diagnosis results. In addition, after training, the diagnostic accuracy of the proposed method can reach 99.54%.

A deep learning model for accurate and robust internet traffic classification

Network traffic classification is significant due to the fast growth of the number of internet users. The traditional way of classifying the large number of traffic generated by these users is becoming less effective. Therefore, many researchers made a network traffic classifier based on deep learning. However, those classifiers do not provide far better results and perform poorly when dealing with encrypted information. This paper tries to approach highly accurate and robust results in both encrypted and unencrypted networks by using machine learning algorithms. The algorithm used is the convolutional neural network (CNN). The performance of the proposed CNN is compared with that of the classical LeNet-5 network. Experimental results show that the classifier based on the proposed CNN performed better when dealing with both encrypted and unencrypted datasets, achieving a maximum average accuracy of 83.55%. Moreover, it is not sensitive to hyper-parameter choices, indicating its superiority in robustness. Compared with traditional network classifiers, the network classifier based on CNN can improve accuracy and improve stability.

Deep Learning-Based Composite Fault Diagnosis

As the core component of many large machinery, rotating equipment occupies a high proportion in industrial manufacturing. Intelligent fault diagnosis for them is becoming the core focus of Industry 4.0 and its evolution. To address the problem of poor fault classification accuracy of rotating equipment, this paper proposes a new solution to deploy a composite fault diagnosis model at the front end to achieve low power consumption, continuous automatic identification and monitoring of equipment faults. We creatively combined the idea of speech recognition to solve the problem that traditional feature extraction methods are insufficient in handling composite fault diagnosis when dealing with fault waveform files, and used Fbank speech feature extraction idea to extract fault features, thus simplifying the complex fault diagnosis problem into an image classification problem. The generated composite fault Fbank feature data set is significantly different in visual observation. Therefore, we further designed the LeNet-F network based on the LeNet network for composite fault diagnosis. The model was validated on the K210 development board to achieve an accuracy of 97.41% for hybrid fault classification at a model size of 2.4MB. This makes it possible to apply the deep learning method to the front end to directly complete the composite fault diagnosis.

Fast Localization and High Accuracy Recognition of Tire Surface Embossed Characters Based on CNN

To solve the problem of recognizing artificial tire-side pressure printing characters with low efficiency and high labor intensity, we propose a CNN-based method for tire surface character recognition. In the image pre-processing, the SSR algorithm is improved to enhance the contrast of characters, and the Normalized Cross Correlation template matching algorithm based on pyramid acceleration is proposed to quickly locate the “DOT” characters and segment them. The improved LeNet-5 network structure is used to recognize characters, and a self-built digital sample library is randomly divided according to the ratio of 8:2 to conduct digital recognition experiments. The experimental results show that the recognition accuracy of the training set can reach 95.9%, and the accuracy of the validation set is 99.5%. The accuracy of the testing set is 95.6%, which meets the practical application requirements. Moreover, the whole algorithm only needs to be implemented on a commonly configured CPU, reducing equipment costs.

Deeply Smile Detection Based on Discriminative Features with Modified LeNet-5 Network

Deeply Smile Detection Based on Discriminative Features with Modified LeNet-5 Network

Adaptive Dimensional Gaussian Mutation of PSO-Optimized Convolutional Neural Network Hyperparameters

The configuration of the hyperparameters in convolutional neural networks (CNN) is crucial for determining their performance. However, traditional methods for hyperparameter configuration, such as grid searches and random searches, are time consuming and labor intensive. The optimization of CNN hyperparameters is a complex problem involving multiple local optima that poses a challenge for traditional particle swarm optimization (PSO) algorithms, which are prone to getting stuck in the local optima and achieving suboptimal results. To address the above issues, we proposed an adaptive dimensional Gaussian mutation PSO (ADGMPSO) to efficiently select the optimal hyperparameter configurations. The ADGMPSO algorithm utilized a cat chaos initialization strategy to generate an initial population with a more uniform distribution. It combined the sine-based inertia weights and an asynchronous change learning factor strategy to balance the global exploration and local exploitation capabilities. Finally, an elite particle adaptive dimensional Gaussian mutation strategy was proposed to improve the population diversity and convergence accuracy at the different stages of evolution. The performance of the proposed algorithm was compared to five other evolutionary algorithms, including PSO, BOA, WOA, SSA, and GWO, on ten benchmark test functions, and the results demonstrated the superiority of the proposed algorithm in terms of the optimal value, mean value, and standard deviation. The ADGMPSO algorithm was then applied to the hyperparameter optimization for the LeNet-5 and ResNet-18 network models. The results on the MNIST and CIFAR10 datasets showed that the proposed algorithm achieved a higher accuracy and generalization ability than the other optimization algorithms, such as PSO-CNN, LDWPSO-CNN, and GA-CNN.

An efficient implementation of traffic signs recognition system using CNN

In the modern world, road traffic signs are vital for drivers safety. In fact, multistep traffic forecasting on road networks can help to avoid many problems on the streets. In this context, there are several methods which allow to achieve excellent results in the field of traffic signs recognition. Recently, Deep Convolutional Neural Network (CNN) have achieved excellent results in this area.In this paper, CNN is used to develop a Traffic and Road Sign recognition system. The performance of the proposed architecture is measured using a novel dataset, namely the Tunisian traffic signs dataset. In addition, we minimize the number of layers in the LeNet network, lowering the number of parameters in the network to accelerate the computation. Our architecture was used with varying parameters in order to achieve the best recognition rates in uncontrolled environment including weather conditions, complex background, variable illumination, and sign color fading. The Experimental results show that the proposed CNN architecture achieved a significant accuracy, thus higher than those achieved in similar previous studies.

Visual Object Tracking Based on Modified LeNet-5 and RCCF

The field of object tracking has recently made significant progress. Particularly, the performance results in both deep learning and correlation filters, based trackers achieved effective tracking performance. Moreover, there are still some difficulties with object tracking for example illumination and deformation (DEF). The precision and accuracy of tracking algorithms suffer from the effects of such occurrences. For this situation, finding a solution is important. This research proposes a new tracking algorithm to handle this problem. The features are extracted by using Modified LeNet-5, and the precision and accuracy are improved by developing the Real-Time Cross-modality Correlation Filtering method (RCCF). In Modified LeNet-5, the visual tracking performance is improved by adjusting the number and size of the convolution kernels in the pooling and convolution layers. The high-level, middle-level, and handcraft features are extracted from the modified LeNet-5 network. The handcraft features are used to determine the specific location of the target because the handcraft features contain more spatial information regarding the visual object. The LeNet features are more suitable for a target appearance change in object tracking. Extensive experiments were conducted by the Object Tracking Benchmarking (OTB) databases like OTB50 and OTB100. The experimental results reveal that the proposed tracker outperforms other state-of-the-art trackers under different problems. The experimental simulation is carried out in python. The overall success rate and precision of the proposed algorithm are 93.8% and 92.5%. The average running frame rate reaches 42 frames per second, which can meet the real-time requirements.

Automatic body condition scoring for dairy cows based on efficient net and convex hull features of point clouds

Body condition score (BCS) is an important indicator to evaluate the nutritional and health status of perinatal dairy cows. The insufficient accuracy of the automatic body condition score is the core bottleneck restricting the popularization and application of this technology. To further improve the accuracy of automatic scoring of cow body condition, a 3D structure feature map based on convex hull distance of point clouds was constructed and used as the input of a two-level model based on EfficientNet network, which realized the improvement of recognition accuracy within the error of 0.25 in the automatic scoring of cow body condition. First, the depth image of the cow’s back was preprocessed to obtain the point cloud from the hook bone to the needle bone region containing the main body condition information. Second, the point cloud was voxelized and convex hulled, and the distance between each peripheral voxel and the nearest convex hull surface was calculated and projected to the X-Y plane to obtain the structural feature map. Then, a two-level classification model based on the EfficientNet network was constructed, and the whale optimization algorithm (WOA) was used to optimize the network scaling coefficient in the model. Finally, 5119 depth images of 77 cows were used to train and test the model, and the dataset ratio was 1:1. The results showed that when the BCS value was in the range of 2.25 to 3.75 and the two-class classification limit was 2.75, the proportions of images with recognition errors less than 0.25 and 0.50 were 91.21% and 97.60%, respectively, and the average recognition speed was 3.446 frames/s. The recognition effect was better than that of the MobileNet-V2, XceptionNet and LeNet-5 network models. This method could realize the noncontact assessment of individual body conditions of dairy cows in large-scale farms and had the characteristics of high accuracy, strong applicability and low cost.

Stacking ensemble learning models for daily runoff prediction using 1D and 2D CNNs

In recent years, applications of convolutional neural networks (CNNs) to runoff prediction have received some attention due to their excellent feature extraction capabilities. However, existing studies are still limited since merely either 1D or 2D CNNs are developed to predict runoff. In this study, a stacking ensemble learning model for daily runoff prediction based on different types of 1D and 2D CNNs is proposed and applied to the Quinebaug River Basin, Connecticut, USA. The structure of the CNN models is developed with reference to the classic LeNet5 network. Especially, the predictors are reconstructed into 1D vectors and 2D matrices with 10-, 20- and 30-day time steps. Totally 18 member models are constructed through selecting 3 representative 1D and 2D CNN models with 3 time steps. The simple average method (SAM) is used to integrate different CNN member models. The results show that the performance of the same-type SAM based on either 1D or 2D CNNs improves because it can counteract the effects of both positive and negative predicted values by the member models to some extent. Furthermore, the mixed-type SAM models based on both 1D and 2D CNN member models can further improve the prediction accuracy. The optimal model SAM15 consists of two 1D CNNs and four 2D CNNs. Compared with the optimal CNN member models, SAM15 reduces the validation RMSE by about 13% and improves the validation R and NSE by about 3% and 7%, respectively. This study highlights that the proposed stacking ensemble learning model can improve the daily runoff prediction accuracy through integration of the nonlinear fitting ability of 1D and 2D CNN member models.

Functional Criticality Analysis of Structural Faults in AI Accelerators

The ubiquitous application of deep neural networks (DNNs) has led to a rise in demand for artificial intelligence (AI) accelerators. For example, the tensor processing unit from Google–based on a systolic array–and its variants are of considerable interest for DNN inferencing using AI accelerators. This article studies the problem of classifying structural faults in such an accelerator based on their functional criticality. We first analyze pin-level faults in the processing elements (PEs) of a systolic array. Simulation results for the LeNet network with 8-bit fixed-point, 16-bit floating-point (FP), and 32-bit FP data paths applied to the MNIST dataset show that over 93% of the pin-level structural faults in a PE are functionally benign. We present a greedy iterative framework for determining the criticality of stuck-at faults in a PE netlist and analyze the limitations of criticality analysis methods based on repeated fault simulations. We next present a scalable two-tier machine-learning (ML)-based method to assess the functional criticality of stuck-at faults in a computationally efficient manner. We address the problem of minimizing misclassification by utilizing generative adversarial networks (GANs). Two-tier ML/GAN-based criticality assessment leads to less than 1% test escapes during functional criticality evaluation of structural faults.

Research on emotional semantic retrieval of attention mechanism oriented to audio-visual synesthesia

Digital video is widely used to record people’s daily lives and share people’s moods, but few researchers have conducted research on the consistency of emotional expression between short videos and music. In order to be able to match the appropriate background music to the short video image autonomously and efficiently, the paper analyzed the emotional connection between the two from the audio-visual synesthesia. First, emotional semantics was used as a bridge to connect video data and music data, and a video-music synesthesia data set based on semantic words was constructed. Then, an attention mechanism was incorporated to better extract key features in video images. In the extraction of music features, an improved lenet5 network was used, and the optimal network parameters were determined through experiments. Finally, the two types of features were fused and the mutual retrieval between video and music was performed. In order to compare the performance of different models, different CNN models were calculated in the processing of video images, including VGG16, VGG19, AlexNet and GoogleNet, and the attention mechanism was added to each network for calculation to compare its retrieval accuracy. In the processing of music data, different CNN algorithms were also used for comparative experiments, and networks with different layers were used to determine the optimal results. The experimental results show that the audiovisual synesthesia retrieval model based on emotion can effectively measure the emotional similarity between video images and music, and the method of the paper can produce a good match between them. The research method of the paper is the exploration of computer synesthetic intelligence, which can stimulate the creative inspiration of image and music creative designers. While enhancing the emotional experience of digital products, it also improves the efficiency and quality of development.

A Hybrid Improved Neural Networks Algorithm Based on L2 and Dropout Regularization

Small samples are prone to overfitting in the neural network training process. This paper proposes an optimization approach based on L2 and dropout regularization called a hybrid improved neural network algorithm to overcome this issue. The proposed model was evaluated based on the Modified National Institute of Standards and Technology (MNIST, grayscale-28 × 28 × 1) and Canadian Institute for Advanced Research 10 (CIFAR10, RGB - 32 × 32 × 3) as the training data sets and data applied to the LeNet-5 and Autoencoder neural network architectures. The evaluation is conducted based on cross-validation; the result of the model prediction is used as the final measure to evaluate the quality of the model. The results show that the proposed hybrid algorithm can perform more effectively, avoid overfitting, improve the accuracy of network model prediction in classification tasks, and reduce the reconstruction error in the unsupervised domain. In addition, employing the proposed algorithm without increasing the time complexity can reduce the effect of noisy data and bias and improve the training time of neural network models. Quantitative and qualitative experimental results show that the accuracy of using the proposed algorithm in this paper with the MNIST test set has an improvement of 2.3% and 0.9% compared to L2 regularization and dropout regularization, respectively, and based on the CIFAR10 data set, the accuracy improvement of 0.92% compared with L2 regularization and 1.31% concerning dropout regularization. The reconstruction error of using the proposed algorithm in this paper with the MNIST data set has an improvement of 0.00174 and 0.00398 compared to L2 regularization and dropout regularization, respectively, and based on the CIFAR10 data set, the accuracy improvement of 0.00078 compared with L2 regularization and 0.00174 concerning dropout regularization.

Research on Detection and Recognition of Traffic Signs Based on Convolutional Neural Networks

The road traffic sign detection and identification (TSDR) system is a subsystem of the advanced driver assistance system. It has received extensive attention from domestic and foreign researchers. The TSDR system is mainly composed of two parts: the detection of traffic signs and the identification and output of traffic signs. This paper focuses on the research of TSDR system and proposes a method based on faster R-CNN combined with part of the low-level feature map. The lightweight model MobileNet is used to finish the traffic sign detection. Experiments show that this method has a certain effect on the detection of small-scale traffic signs, and the detection speed is fast. The LeNet5 and Inception V3 network model are used for traffic sign recognition. The model is optimized by adjusting parameters such as the learning rate and batch size. It shows that for LeNet5, when the learning rate = 0.001, the recognition rate can reach 92.5%. For the Inception V3 model, when the learning rate is 0.005, it has a higher recognition rate than the LeNet5 model by 4.9 percentage points.

Rapid detection of Penaeus vannamei diseases via an improved LeNet

Shrimp disease is a greatly important factor in the culture of Penaeus vannamei, the shrimp species with the highest yield in the world aquaculture industry. Hepatopancreatic necrosis disease (HPND, 37 %), red body disease (RBD, 26 %), and whitish muscle disease (WMD, 18 %) were the most common Penaeus vannamei diseases, all of which are usually classified and identified by two kinds of detection (manual detection and germs purifying method). Most of these detections suffer from the class low accuracy, too complex, or too costly. In this study, we tackle this situation with an improved LeNet, which includes modifying model parameters and computational methods. More particularly, this study proposes a convolutional neural networks (CNN) model that is based on LeNet network framework and can reduce parameters and accelerate calculation. To offer improvements in classification and identification, we increase the number of feature maps. Meanwhile, to firstly take denoise and then strengthen characteristic in pretreatment, HSV color space conversion and Gaussian noise with a level of 20 are led into. We conclude that the model generates the precision at about 96.1 % when the weight parameter learning rate is 0.002 and the number of iterations is 120 after being trained and validated. The study has made tremendous progress in the rapid detection of Penaeus vannamei diseases by providing an effective technological path and suggesting the possibility of realizing early disease warnings in future works.