A Lightweight Convolutional Neural Network with Hierarchical Multi-Scale Feature Fusion for Image Classification

Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware resources. To address this issue, the MobileNetV1 network was developed, which employs depthwise convolution to reduce network complexity. MobileNetV1 employs a stride of 2 in several convolutional layers to decrease the spatial resolution of feature maps, thereby lowering computational costs. However, this stride setting can lead to a loss of spatial information, particularly affecting the detection and representation of smaller objects or finer details in images. To maintain the trade-off between complexity and model performance, a lightweight convolutional neural network with hierarchical multi-scale feature fusion based on the MobileNetV1 network is proposed. The network consists of two main subnetworks. The first subnetwork uses a depthwise dilated separable convolution (DDSC) layer to learn imaging features with fewer parameters, which results in a lightweight and computationally inexpensive network. Furthermore, depthwise dilated convolution in DDSC layer effectively expands the field of view of filters, allowing them to incorporate a larger context. The second subnetwork is a hierarchical multi-scale feature fusion (HMFF) module that uses parallel multi-resolution branches architecture to process the input feature map in order to extract the multi-scale feature information of the input image. Experimental results on the CIFAR-10, Malaria, and KvasirV1 datasets demonstrate that the proposed method is efficient, reducing the network parameters and computational cost by 65.02% and 39.78%, respectively, while maintaining the network performance compared to the MobileNetV1 baseline.

A COVID-19 Detection Model Based on Convolutional Neural Network and Residual Learning

Amodel that can obtain rapid and accurate detection of coronavirus disease 2019(COVID-19)plays a significant role in treating and preventing the spread of disease transmission.However,designing such amodel that can balance the detection accuracy andweight parameters ofmemorywell to deploy a mobile device is challenging.Taking this point into account,this paper fuses the convolutional neural network and residual learning operations to build a multi-class classification model,which improves COVID-19 pneumonia detection performance and keeps a trade-off between the weight parameters and accuracy.The convolutional neural network can extract the COVID-19 feature information by repeated convolutional operations.The residual learning operations alleviate the gradient problems caused by stacking convolutional layers and enhance the ability of feature extraction.The ability further enables the proposed model to acquire effective feature information at a lowcost,which canmake ourmodel keep smallweight parameters.Extensive validation and comparison with other models of COVID-19 pneumonia detection on the well-known COVIDx dataset show that(1)the sensitivity of COVID-19 pneumonia detection is improved from 88.2%(non-COVID-19)and 77.5%(COVID-19)to 95.3%(non-COVID-19)and 96.5%(COVID-19),respectively.The positive predictive value is also respectively increased from72.8%(non-COVID-19)and 89.0%(COVID-19)to 88.8%(non-COVID-19)and 95.1%(COVID-19).(2)Compared with the weight parameters of the COVIDNet-small network,the value of the proposed model is 13 M,which is slightly higher than that(11.37 M)of the COVIDNet-small network.But,the corresponding accuracy is improved from 85.2%to 93.0%.The above results illustrate the proposed model can gain an efficient balance between accuracy and weight parameters.

Rockburst Intensity Grade Prediction Model Based on Batch Gradient Descent and Multi-Scale Residual Deep Neural Network

Rockburst is a phenomenon in which free surfaces are formed during excavation,which subsequently causes the sudden release of energy in the construction of mines and tunnels.Light rockburst only peels off rock slices without ejection,while severe rockburst causes casualties and property loss.The frequency and degree of rockburst damage increases with the excavation depth.Moreover,rockburst is the leading engineering geological hazard in the excavation process,and thus the prediction of its intensity grade is of great significance to the development of geotechnical engineering.Therefore,the prediction of rockburst intensity grade is one problem that needs to be solved urgently.By comprehensively considering the occurrence mechanism of rockburst,this paper selects the stress index(σθ/σc),brittleness index(σ_(c)/σ_(t)),and rock elastic energy index(Wet)as the rockburst evaluation indexes through the Spearman coefficient method.This overcomes the low accuracy problem of a single evaluation index prediction method.Following this,the BGD-MSR-DNN rockburst intensity grade prediction model based on batch gradient descent and a multi-scale residual deep neural network is proposed.The batch gradient descent(BGD)module is used to replace the gradient descent algorithm,which effectively improves the efficiency of the network and reduces the model training time.Moreover,the multi-scale residual(MSR)module solves the problem of network degradation when there are too many hidden layers of the deep neural network(DNN),thus improving the model prediction accuracy.The experimental results reveal the BGDMSR-DNN model accuracy to reach 97.1%,outperforming other comparable models.Finally,actual projects such as Qinling Tunnel and Daxiangling Tunnel,reached an accuracy of 100%.The model can be applied in mines and tunnel engineering to realize the accurate and rapid prediction of rockburst intensity grade.

Automatic modulation recognition of radiation source signals based on two-dimensional data matrix and improved residual neural network

Automatic modulation recognition(AMR)of radiation source signals is a research focus in the field of cognitive radio.However,the AMR of radiation source signals at low SNRs still faces a great challenge.Therefore,the AMR method of radiation source signals based on two-dimensional data matrix and improved residual neural network is proposed in this paper.First,the time series of the radiation source signals are reconstructed into two-dimensional data matrix,which greatly simplifies the signal preprocessing process.Second,the depthwise convolution and large-size convolutional kernels based residual neural network(DLRNet)is proposed to improve the feature extraction capability of the AMR model.Finally,the model performs feature extraction and classification on the two-dimensional data matrix to obtain the recognition vector that represents the signal modulation type.Theoretical analysis and simulation results show that the AMR method based on two-dimensional data matrix and improved residual network can significantly improve the accuracy of the AMR method.The recognition accuracy of the proposed method maintains a high level greater than 90% even at -14 dB SNR.

Multi-Classification of Polyps in Colonoscopy Images Based on an Improved Deep Convolutional Neural Network

Achieving accurate classification of colorectal polyps during colonoscopy can avoid unnecessary endoscopic biopsy or resection.This study aimed to develop a deep learning model that can automatically classify colorectal polyps histologically on white-light and narrow-band imaging(NBI)colonoscopy images based on World Health Organization(WHO)and Workgroup serrAted polypS and Polyposis(WASP)classification criteria for colorectal polyps.White-light and NBI colonoscopy images of colorectal polyps exhibiting pathological results were firstly collected and classified into four categories:conventional adenoma,hyperplastic polyp,sessile serrated adenoma/polyp(SSAP)and normal,among which conventional adenoma could be further divided into three sub-categories of tubular adenoma,villous adenoma and villioustublar adenoma,subsequently the images were re-classified into six categories.In this paper,we proposed a novel convolutional neural network termed Polyp-DedNet for the four-and six-category classification tasks of colorectal polyps.Based on the existing classification network ResNet50,Polyp-DedNet adopted dilated convolution to retain more high-dimensional spatial information and an Efficient Channel Attention(ECA)module to improve the classification performance further.To eliminate gridding artifacts caused by dilated convolutions,traditional convolutional layers were used instead of the max pooling layer,and two convolutional layers with progressively decreasing dilation were added at the end of the network.Due to the inevitable imbalance of medical image data,a regularization method DropBlock and a Class-Balanced(CB)Loss were performed to prevent network overfitting.Furthermore,the 5-fold cross-validation was adopted to estimate the performance of Polyp-DedNet for the multi-classification task of colorectal polyps.Mean accuracies of the proposed Polyp-DedNet for the four-and six-category classifications of colorectal polyps were 89.91%±0.92%and 85.13%±1.10%,respectively.The metrics of precision,recall and F1-score were also improved by 1%∼2%compared to the baseline ResNet50.The proposed Polyp-DedNet presented state-of-the-art performance for colorectal polyp classifying on white-light and NBI colonoscopy images,highlighting its considerable potential as an AI-assistant system for accurate colorectal polyp diagnosis in colonoscopy.

Importance of Adaptive Photometric Augmentation for Different Convolutional Neural Network

Existing segmentation and augmentation techniques on convolutional neural network(CNN)has produced remarkable progress in object detection.However,the nominal accuracy and performance might be downturned with the photometric variation of images that are directly ignored in the training process,along with the context of the individual CNN algorithm.In this paper,we investigate the effect of a photometric variation like brightness and sharpness on different CNN.We observe that random augmentation of images weakens the performance unless the augmentation combines the weak limits of photometric variation.Our approach has been justified by the experimental result obtained from the PASCAL VOC 2007 dataset,with object detection CNN algorithms such as YOLOv3(You Only Look Once),Faster R-CNN(Region-based CNN),and SSD(Single Shot Multibox Detector).Each CNN model shows performance loss for varying sharpness and brightness,ranging between−80%to 80%.It was further shown that compared to random augmentation,the augmented dataset with weak photometric changes delivered high performance,but the photometric augmentation range differs for each model.Concurrently,we discuss some research questions that benefit the direction of the study.The results prove the importance of adaptive augmentation for individual CNN model,subjecting towards the robustness of object detection.

Pedestrian attribute classification with multi-scale and multi-label convolutional neural networks

Pedestrian attribute classification from a pedestrian image captured in surveillance scenarios is challenging due to diverse clothing appearances,varied poses and different camera views. A multiscale and multi-label convolutional neural network( MSMLCNN) is proposed to predict multiple pedestrian attributes simultaneously. The pedestrian attribute classification problem is firstly transformed into a multi-label problem including multiple binary attributes needed to be classified. Then,the multi-label problem is solved by fully connecting all binary attributes to multi-scale features with logistic regression functions. Moreover,the multi-scale features are obtained by concatenating those featured maps produced from multiple pooling layers of the MSMLCNN at different scales. Extensive experiment results show that the proposed MSMLCNN outperforms state-of-the-art pedestrian attribute classification methods with a large margin.

Object Grasping Detection Based on Residual Convolutional Neural Network

Robotic grasps play an important role in the service and industrial fields,and the robotic arm can grasp the object properly depends on the accuracy of the grasping detection result.In order to predict grasping detection positions for known or unknown objects by a modular robotic system,a convolutional neural network(CNN)with the residual block is proposed,which can be used to generate accurate grasping detection for input images of the scene.The proposed model architecture was trained on the standard Cornell grasp dataset and evaluated on the test dataset.Moreover,it was evaluated on different types of household objects and cluttered multi-objects.On the Cornell grasp dataset,the accuracy of the model on image-wise splitting detection and object-wise splitting detection achieved 95.5%and 93.6%,respectively.Further,the real detection time per image was 109 ms.The experimental results show that the model can quickly detect the grasping positions of a single object or multiple objects in image pixels in real time,and it keeps good stability and robustness.

Vehicle Plate Number Localization Using Memetic Algorithms and Convolutional Neural Networks

This paper introduces the third enhanced version of a genetic algorithm-based technique to allow fast and accurate detection of vehicle plate numbers(VPLN)in challenging image datasets.Since binarization of the input image is the most important and difficult step in the detection of VPLN,a hybrid technique is introduced that fuses the outputs of three fast techniques into a pool of connected components objects(CCO)and hence enriches the solution space with more solution candidates.Due to the combination of the outputs of the three binarization techniques,many CCOs are produced into the output pool from which one or more sequences are to be selected as candidate solutions.The pool is filtered and submitted to a new memetic algorithm to select the best fit sequence of CCOs based on an objective distance between the tested sequence and the defined geometrical relationship matrix that represents the layout of the VPLN symbols inside the concerned plate prototype.Using any of the previous versions will give moderate results but with very low speed.Hence,a new local search is added as a memetic operator to increase the fitness of the best chromosomes based on the linear arrangement of the license plate symbols.The memetic operator speeds up the convergence to the best solution and hence compensates for the overhead of the used hybrid binarization techniques and allows for real-time detection especially after using GPUs in implementing most of the used techniques.Also,a deep convolutional network is used to detect false positives to prevent fake detection of non-plate text or similar patterns.Various image samples with a wide range of scale,orientation,and illumination conditions have been experimented with to verify the effect of the new improvements.Encouraging results with 97.55%detection precision have been reported using the recent challenging public Chinese City Parking Dataset(CCPD)outperforming the author of the dataset by 3.05%and the state-of-the-art technique by 1.45%.

MSSTNet:Multi-scale facial videos pulse extraction network based on separable spatiotemporal convolution and dimension separable attention

Background The use of remote photoplethysmography(rPPG)to estimate blood volume pulse in a noncontact manner has been an active research topic in recent years.Existing methods are primarily based on a singlescale region of interest(ROI).However,some noise signals that are not easily separated in a single-scale space can be easily separated in a multi-scale space.Also,existing spatiotemporal networks mainly focus on local spatiotemporal information and do not emphasize temporal information,which is crucial in pulse extraction problems,resulting in insufficient spatiotemporal feature modelling.Methods Here,we propose a multi-scale facial video pulse extraction network based on separable spatiotemporal convolution(SSTC)and dimension separable attention(DSAT).First,to solve the problem of a single-scale ROI,we constructed a multi-scale feature space for initial signal separation.Second,SSTC and DSAT were designed for efficient spatiotemporal correlation modeling,which increased the information interaction between the long-span time and space dimensions;this placed more emphasis on temporal features.Results The signal-to-noise ratio(SNR)of the proposed network reached 9.58dB on the PURE dataset and 6.77dB on the UBFC-rPPG dataset,outperforming state-of-the-art algorithms.Conclusions The results showed that fusing multi-scale signals yielded better results than methods based on only single-scale signals.The proposed SSTC and dimension-separable attention mechanism will contribute to more accurate pulse signal extraction.

Lightweight Image Super-Resolution via Weighted Multi-Scale Residual Network

The tradeoff between efficiency and model size of the convolutional neural network(CNN)is an essential issue for applications of CNN-based algorithms to diverse real-world tasks.Although deep learning-based methods have achieved significant improvements in image super-resolution(SR),current CNNbased techniques mainly contain massive parameters and a high computational complexity,limiting their practical applications.In this paper,we present a fast and lightweight framework,named weighted multi-scale residual network(WMRN),for a better tradeoff between SR performance and computational efficiency.With the modified residual structure,depthwise separable convolutions(DS Convs)are employed to improve convolutional operations’efficiency.Furthermore,several weighted multi-scale residual blocks(WMRBs)are stacked to enhance the multi-scale representation capability.In the reconstruction subnetwork,a group of Conv layers are introduced to filter feature maps to reconstruct the final high-quality image.Extensive experiments were conducted to evaluate the proposed model,and the comparative results with several state-of-the-art algorithms demonstrate the effectiveness of WMRN.

Multi-Scale Convolutional Gated Recurrent Unit Networks for Tool Wear Prediction in Smart Manufacturing

As an integrated application of modern information technologies and artificial intelligence,Prognostic and Health Management(PHM)is important for machine health monitoring.Prediction of tool wear is one of the symbolic applications of PHM technology in modern manufacturing systems and industry.In this paper,a multi-scale Convolutional Gated Recurrent Unit network(MCGRU)is proposed to address raw sensory data for tool wear prediction.At the bottom of MCGRU,six parallel and independent branches with different kernel sizes are designed to form a multi-scale convolutional neural network,which augments the adaptability to features of different time scales.These features of different scales extracted from raw data are then fed into a Deep Gated Recurrent Unit network to capture long-term dependencies and learn significant representations.At the top of the MCGRU,a fully connected layer and a regression layer are built for cutting tool wear prediction.Two case studies are performed to verify the capability and effectiveness of the proposed MCGRU network and results show that MCGRU outperforms several state-of-the-art baseline models.

融合Residual Network-50残差块与卷积注意力模块的地震断层自动识别

传统的断层识别是由地质解释人员以人工标记的方式进行检测,不仅耗时长、效率低,且识别结果存在一定的人为误差。为解决以上问题,提高断层识别的精度,提出了一种基于深度学习的断层识别方法,利用注意力机制聚焦目标特征的能力,在U-Net网络的解码层引入了卷积注意力模块(Convolutional Block Attention Module,CBAM),在编码层引入了ResNet-50残差块,建立基于卷积神经网络(Convolutional Neural Networks,CNN)的断层识别方法(Res-CBAM-UNet)。将合成地震数据与相应的断层标签进行数据增强操作,新生成的训练数据集作为输入对网络模型进行训练,以提高模型的泛化能力。随后将该模型与CBAM-UNet、ResNet34-UNet和ResNet50-UNet网络进行对比分析,利用实际工区地震数据进行测试。结果表明,设计的Res-CBAM-UNet网络对断层具有较好的识别效果,且识别出的断层连续性好,计算效率高。

Image recognition and empirical application of desert plant species based on convolutional neural network

In recent years,deep convolution neural network has exhibited excellent performance in computer vision and has a far-reaching impact.Traditional plant taxonomic identification requires high expertise,which is time-consuming.Most nature reserves have problems such as incomplete species surveys,inaccurate taxonomic identification,and untimely updating of status data.Simple and accurate recognition of plant images can be achieved by applying convolutional neural network technology to explore the best network model.Taking 24 typical desert plant species that are widely distributed in the nature reserves in Xinjiang Uygur Autonomous Region of China as the research objects,this study established an image database and select the optimal network model for the image recognition of desert plant species to provide decision support for fine management in the nature reserves in Xinjiang,such as species investigation and monitoring,by using deep learning.Since desert plant species were not included in the public dataset,the images used in this study were mainly obtained through field shooting and downloaded from the Plant Photo Bank of China(PPBC).After the sorting process and statistical analysis,a total of 2331 plant images were finally collected(2071 images from field collection and 260 images from the PPBC),including 24 plant species belonging to 14 families and 22 genera.A large number of numerical experiments were also carried out to compare a series of 37 convolutional neural network models with good performance,from different perspectives,to find the optimal network model that is most suitable for the image recognition of desert plant species in Xinjiang.The results revealed 24 models with a recognition Accuracy,of greater than 70.000%.Among which,Residual Network X_8GF(RegNetX_8GF)performs the best,with Accuracy,Precision,Recall,and F1(which refers to the harmonic mean of the Precision and Recall values)values of 78.33%,77.65%,69.55%,and 71.26%,respectively.Considering the demand factors of hardware equipment and inference time,Mobile NetworkV2 achieves the best balance among the Accuracy,the number of parameters and the number of floating-point operations.The number of parameters for Mobile Network V2(MobileNetV2)is 1/16 of RegNetX_8GF,and the number of floating-point operations is 1/24.Our findings can facilitate efficient decision-making for the management of species survey,cataloging,inspection,and monitoring in the nature reserves in Xinjiang,providing a scientific basis for the protection and utilization of natural plant resources.

Defect Detection Algorithm of Patterned Fabrics Based on Convolutional Neural Network

The background pattern of patterned fabrics is complex,which has a great interference in the extraction of defect features.Traditional machine vision algorithms rely on artificially designed features,which are greatly affected by background patterns and are difficult to effectively extract flaw features.Therefore,a convolutional neural network(CNN)with automatic feature extraction is proposed.On the basis of the two-stage detection model Faster R-CNN,Resnet-50 is used as the backbone network,and the problem of flaws with extreme aspect ratio is solved by improving the initialization algorithm of the prior frame aspect ratio,and the improved multi-scale model is designed to improve detection of small defects.The cascade R-CNN is introduced to improve the accuracy of defect detection,and the online hard example mining(OHEM)algorithm is used to strengthen the learning of hard samples to reduce the interference of complex backgrounds on the defect detection of patterned fabrics,and construct the focal loss as a loss function to reduce the impact of sample imbalance.In order to verify the effectiveness of the improved algorithm,a defect detection comparison experiment was set up.The experimental results show that the accuracy of the defect detection algorithm of patterned fabrics in this paper can reach 95.7%,and it can accurately locate the defect location and meet the actual needs of the factory.

A Spectral Convolutional Neural Network Model Based on Adaptive Fick’s Law for Hyperspectral Image Classification

Hyperspectral image classification stands as a pivotal task within the field of remote sensing,yet achieving highprecision classification remains a significant challenge.In response to this challenge,a Spectral Convolutional Neural Network model based on Adaptive Fick’s Law Algorithm(AFLA-SCNN)is proposed.The Adaptive Fick’s Law Algorithm(AFLA)constitutes a novel metaheuristic algorithm introduced herein,encompassing three new strategies:Adaptive weight factor,Gaussian mutation,and probability update policy.With adaptive weight factor,the algorithmcan adjust theweights according to the change in the number of iterations to improve the performance of the algorithm.Gaussianmutation helps the algorithm avoid falling into local optimal solutions and improves the searchability of the algorithm.The probability update strategy helps to improve the exploitability and adaptability of the algorithm.Within the AFLA-SCNN model,AFLA is employed to optimize two hyperparameters in the SCNN model,namely,“numEpochs”and“miniBatchSize”,to attain their optimal values.AFLA’s performance is initially validated across 28 functions in 10D,30D,and 50D for CEC2013 and 29 functions in 10D,30D,and 50D for CEC2017.Experimental results indicate AFLA’s marked performance superiority over nine other prominent optimization algorithms.Subsequently,the AFLA-SCNN model was compared with the Spectral Convolutional Neural Network model based on Fick’s Law Algorithm(FLA-SCNN),Spectral Convolutional Neural Network model based on Harris Hawks Optimization(HHO-SCNN),Spectral Convolutional Neural Network model based onDifferential Evolution(DE-SCNN),SpectralConvolutionalNeuralNetwork(SCNN)model,and SupportVector Machines(SVM)model using the Indian Pines dataset and PaviaUniversity dataset.The experimental results show that the AFLA-SCNN model outperforms other models in terms of Accuracy,Precision,Recall,and F1-score on Indian Pines and Pavia University.Among them,the Accuracy of the AFLA-SCNN model on Indian Pines reached 99.875%,and the Accuracy on PaviaUniversity reached 98.022%.In conclusion,our proposed AFLA-SCNN model is deemed to significantly enhance the precision of hyperspectral image classification.

A Novel Forgery Detection in Image Frames of the Videos Using Enhanced Convolutional Neural Network in Face Images

Different devices in the recent era generated a vast amount of digital video.Generally,it has been seen in recent years that people are forging the video to use it as proof of evidence in the court of justice.Many kinds of researches on forensic detection have been presented,and it provides less accuracy.This paper proposed a novel forgery detection technique in image frames of the videos using enhanced Convolutional Neural Network(CNN).In the initial stage,the input video is taken as of the dataset and then converts the videos into image frames.Next,perform pre-sampling using the Adaptive Rood Pattern Search(ARPS)algorithm intended for reducing the useless frames.In the next stage,perform preprocessing for enhancing the image frames.Then,face detection is done as of the image utilizing the Viola-Jones algorithm.Finally,the improved Crow Search Algorithm(ICSA)has been used to select the extorted features and inputted to the Enhanced Convolutional Neural Network(ECNN)classifier for detecting the forged image frames.The experimental outcome of the proposed system has achieved 97.21%accuracy compared to other existing methods.

Multimodal 3D Convolutional Neural Networks for Classification of Brain Disease Using Structural MR and FDG-PET Images

The classification and identification of brain diseases with multimodal information have attracted increasing attention in the domain of computer-aided. Compared with traditional method which use single modal feature information, multiple modal information fusion can classify and diagnose brain diseases more comprehensively and accurately in patient subjects. Existing multimodal methods require manual extraction of features or additional personal information, which consumes a lot of manual work. Furthermore, the difference between different modal images along with different manual feature extraction make it difficult for models to learn the optimal solution. In this paper, we propose a multimodal 3D convolutional neural networks framework for classification of brain disease diagnosis using MR images data and PET images data of subjects. We demonstrate the performance of the proposed approach for classification of Alzheimer’s disease (AD) versus mild cognitive impairment (MCI) and normal controls (NC) on the Alzheimer’s Disease National Initiative (ADNI) data set of 3D structural MRI brain scans and FDG-PET images. Experimental results show that the performance of the proposed method for AD vs. NC, MCI vs. NC are 93.55% and 78.92% accuracy respectively. And the accuracy of the results of AD, MCI and NC 3-classification experiments is 68.86%.

Range estimation of few-shot underwater sound source in shallow water based on transfer learning and residual CNN

Taking the real part and the imaginary part of complex sound pressure of the sound field as features,a transfer learning model is constructed.Based on the pre-training of a large amount of underwater acoustic data in the preselected sea area using the convolutional neural network(CNN),the few-shot underwater acoustic data in the test sea area are retrained to study the underwater sound source ranging problem.The S5 voyage data of SWellEX-96 experiment is used to verify the proposed method,realize the range estimation for the shallow source in the experiment,and compare the range estimation performance of the underwater target sound source of four methods:matched field processing(MFP),generalized regression neural network(GRNN),traditional CNN,and transfer learning.Experimental data processing results show that the transfer learning model based on residual CNN can effectively realize range estimation in few-shot scenes,and the estimation performance is remarkably better than that of other methods.

Identification of tomato leaf diseases using convolutional neural network with multi-scale and feature reuse

Various diseases seriously affect the quality and yield of tomatoes. Fast and accurate identification of disease types is of great significance for the development of smart agriculture. Many Convolution Neural Network (CNN) models have been applied to the identification of tomato leaf diseases and achieved good results. However, some of these are executed at the cost of large calculation time and huge storage space. This study proposed a lightweight CNN model named MFRCNN, which is established by the multi-scale and feature reuse structure rather than simply stacking convolution layer by layer. To examine the model performances, two types of tomato leaf disease datasets were collected. One is the laboratory-based dataset, including one healthy and nine diseases, and the other is the field-based dataset, including five kinds of diseases. Afterward, the proposed MFRCNN and some popular CNN models (AlexNet, SqueezeNet, VGG16, ResNet18, and GoogLeNet) were tested on the two datasets. The results showed that compared to traditional models, the MFRCNN achieved the optimal performance, with an accuracy of 99.01% and 98.75% in laboratory and field datasets, respectively. The MFRCNN not only had the highest accuracy but also had relatively less computing time and few training parameters. Especially in terms of storage space, the MFRCNN model only needs 2.7 MB of space. Therefore, this work provides a novel solution for plant disease diagnosis, which is of great importance for the development of plant disease diagnosis systems on low-performance terminals.