Depth-wise Convolution Research Articles

A Family of Automatic Modulation Classification Models Based on Domain Knowledge for Various Platforms

Identifying the modulation type of radio signals is challenging in both military and civilian applications such as radio monitoring and spectrum allocation. This has become more difficult as the number of signal types increases and the channel environment becomes more complex. Deep learning-based automatic modulation classification (AMC) methods have recently achieved state-of-the-art performance with massive amounts of data. However, existing models struggle to achieve the required level of accuracy, guarantee real-time performance at edge devices, and achieve higher classification performance on high-performance computing platforms when deployed on various platforms. In this paper, we present a family of AMC models based on communication domain knowledge for various computing platforms. The higher-order statistical properties of signals, customized data augmentation methods, and narrowband convolution kernels are the domain knowledge that is specifically employed to the AMC task and neural network backbone. We used separable convolution and depth-wise convolution with very few residual connections to create our lightweight model, which has only 4.61k parameters while maintaining accuracy. On the four different platforms, the classification accuracy and inference time outperformed those of the existing lightweight models. Meanwhile, we use the squeeze-and-excitation attention mechanism, channel shuffle module, and expert feature parallel branch to improve the classification accuracy. On the three most frequently used benchmark datasets, the high-accuracy models achieved state-of-the-art average accuracies of 64.63%, 67.22%, and 65.03%, respectively. Furthermore, we propose a generic framework for evaluating the complexity of deep learning models and use it to comprehensively assess the complexity strengths of the proposed models.

PMA-Net: A parallelly mixed attention network for person re-identification

The objective of person re-identification (Re-ID) is to match unmistakable people across different settings and camera views. Although the use of convolutional neural networks (CNN) has been effective, they lose much significant data and ignore the correlation between the overall and partial features brought about by downsampling operators (e.g., pooling). This compromises the retrieval performance based on the Euclidean distance. We propose a parallelly mixed attention network (PMA-Net) to demonstrate the complementarity of spatial and channel information for Re-ID. First, we integrate self-attention with depthwise convolution in a parallel design, extracting different dimensions of information. Second, we propose two interactive branches to entwine spatial and channel information while extracting features. Then, we add a new additional embedding to incorporate non-visual inputs and reduce the feature bias toward camera variances. Finally, we perform additional maximizing the minimal pairwise angles (MMA) regularization on the features of PMA, which successfully encourages the angular diversity of feature vectors. We confirm the effectiveness of PMA-Net through an extensive set of ablation studies. The results of three benchmarks show the superiority of our model over current mAP/top1 accuracies of 91.3%/96.1% on Market-1501, 83.3%/91.7% on DukeMTMC-ReID, and 66.2%/85.1% on MSMT17.

Local–global lightweight ViT model for mini/micro-LED-chip defect recognition

With the miniaturization of the light-emitting diodes (LEDs), it is a challenge to detect the defective LED-chip with neural networks. Convolution neural networks (CNNs) can capture local information, but the deeper or wider layer with a higher computational cost is required to achieve a stronger model capacity. Additionally, the vision transformer (ViT) is able to extract global information at the cost of large datasets, extensive data augmentation and computational complexity. To solve the aforementioned issues, this paper proposes a local–global lightweight ViT model based on Convmixer for mini/micro-LED-chip defect recognition, which combines the advantages of CNNs and ViT to capture both detailed and global information and is suitable for resource-constrained embedded or mobile devices. A fine-coarse-grained convolutional block is proposed to capture different sizes of features in the spatial dimension with depth-wise convolution. Then, the multi-mode and cross-channel information is extracted by the point-wise convolution. Finally, a position attention module embedded with the long-range relationship is presented to highlight the important patches and captures global information. Besides, the structure of the proposed model is plain so that the detailed information is preserved. Experiments show that the proposed model achieves a higher accuracy of 96.17% with fewer parameters and of 0.038M and a lower computational cost of 0.138G, which outperforms the state-of-the-art.

A COVID-19 medical image classification algorithm based on Transformer

Coronavirus 2019 (COVID-19) is a new acute respiratory disease that has spread rapidly throughout the world. This paper proposes a novel deep learning network based on ResNet-50 merged transformer named RMT-Net. On the backbone of ResNet-50, it uses Transformer to capture long-distance feature information, adopts convolutional neural networks and depth-wise convolution to obtain local features, reduce the computational cost and acceleration the detection process. The RMT-Net includes four stage blocks to realize the feature extraction of different receptive fields. In the first three stages, the global self-attention method is adopted to capture the important feature information and construct the relationship between tokens. In the fourth stage, the residual blocks are used to extract the details of feature. Finally, a global average pooling layer and a fully connected layer perform classification tasks. Training, verification and testing are carried out on self-built datasets. The RMT-Net model is compared with ResNet-50, VGGNet-16, i-CapsNet and MGMADS-3. The experimental results show that the RMT-Net model has a Test_ acc of 97.65% on the X-ray image dataset, 99.12% on the CT image dataset, which both higher than the other four models. The size of RMT-Net model is only 38.5 M, and the detection speed of X-ray image and CT image is 5.46 ms and 4.12 ms per image, respectively. It is proved that the model can detect and classify COVID-19 with higher accuracy and efficiency.

A High-Performance FPGA-Based Depthwise Separable Convolution Accelerator

Depthwise separable convolution (DSC) significantly reduces parameter and floating operations with an acceptable loss of accuracy and has been widely used in various lightweight convolutional neural network (CNN) models. In practical applications, however, DSC accelerators based on graphics processing units (GPUs) cannot fully exploit the performance of DSC and are unsuitable for mobile application scenarios. Moreover, low resource utilization due to idle engines is a common problem in DSC accelerator design. In this paper, a high-performance DSC hardware accelerator based on field-programmable gate arrays (FPGAs) is proposed. A highly reusable and scalable multiplication and accumulation engine is proposed to improve the utilization of computational resources. An efficient convolution algorithm is proposed for depthwise convolution (DWC) and pointwise convolution (PWC), respectively, to reduce the on-chip memory occupancy. Meanwhile, the proposed convolution algorithms achieve partial fusion between PWC and DWC, and improve the off-chip memory access efficiency. To maximise bandwidth utilization and reduce latency when reading feature maps, an address mapping method for off-chip accesses is proposed. The performance of the proposed accelerator is demonstrated by implementing MobileNetV2 on an Intel Arria 10 GX660 FPGA by using Verilog HDL. The experimental results show that the proposed DSC accelerator achieves a performance of 205.1 FPS, 128.8 GFLOPS, and 0.24 GOPS/DSP for input images of size 224×224×3.

Compressing convolutional neural networks with cheap convolutions and online distillation

Visual impairment assistance systems play a vital role in improving the standard of living for visually impaired people (VIP). With the development of deep learning technologies and assistive devices, many assistive technologies for VIP have achieved remarkable success in environmental perception and navigation. In particular, convolutional neural network (CNN)-based models have surpassed the level of human recognition and achieved a strong generalization ability. However, the large memory and computation consumption in CNNs have been one of the main barriers to deploying them into resource-limited systems for visual impairment assistance applications. To this end, most cheap convolutions (e.g., group convolution, depth-wise convolution, and shift convolution) have recently been used for memory and computation reduction but with a specific architecture design. Furthermore, it results in a low discriminability of the compressed networks by directly replacing the standard convolution with these cheap ones. In this paper, we propose to use knowledge distillation to improve the performance of compact student networks with cheap convolutions. In our case, the teacher is a network with the standard convolution, while the student is a simple transformation of the teacher architecture without complicated redesigning. In particular, we introduce a novel online distillation method, which online constructs the teacher network without pre-training and conducts mutual learning between the teacher and student network, to improve the performance of the student model. Extensive experiments demonstrate that the proposed approach achieves superior performance to simultaneously reduce memory and computation overhead of cutting-edge CNNs on different datasets, including CIFAR-10/100 and ImageNet ILSVRC 2012, compared to the previous CNN compression and acceleration methods. The codes are publicly available at https://github.com/EthanZhangYC/OD-cheap-convolution.

LightR-YOLOv5: A compact rotating detector for SARS-CoV-2 antigen-detection rapid diagnostic test results

Nucleic acid testing is currently the golden reference for coronaviruses (SARS-CoV-2) detection, while the SARS-CoV-2 antigen-detection rapid diagnostic tests (RDT) is an important adjunct. RDT can be widely used in the community or regional screening management as self-test tools and may need to be verified by healthcare authorities. However, manual verification of RDT results is a time-consuming task, and existing object detection algorithms usually suffer from high model complexity and computational effort, making them difficult to deploy. We propose LightR-YOLOv5, a compact rotating SARS-CoV-2 antigen-detection RDT results detector. Firstly, we employ an extremely light-weight L-ShuffleNetV2 network as a feature extraction network with a slight reduction in recognition accuracy. Secondly, we combine semantic and texture features in different layers by judiciously combining and employing GSConv, depth-wise convolution, and other modules, and further employ the NAM attention to locate the RDT result detection region. Furthermore, we propose a new data augmentation approach, Single-Copy–Paste, for increasing data samples for the specific task of RDT result detection while achieving a small improvement in model accuracy. Compared with some mainstream rotating object detection networks, the model size of our LightR-YOLOv5 is only 2.03MB, and it is 12.6%, 6.4%, and 7.3% higher in mAP@.5:.95 metrics compared to RetianNet, FCOS, and R3Det, respectively.

DGC‐UWnet: Underwater image enhancement based on computation‐efficient convolution and channel shuffle

AbstractUnderwater image enhancement is receiving increasing attention due to many of the vision research being applied to underwater scenes. To eliminate the impact of complex underwater scenes on imaging, underwater image enhancement algorithm has become an effective solution. However, underwater image enhancement models face a challenge of lightening the model while improving generalizability. Here, DGC‐UWnet is proposed to go for both lightweight and enhancement effect. The proposed model is designed by using depthwise convolution, group convolution and channel shuffle (DGC). Ablation experiment shows that compared with standard convolution, DGC decreases model parameters and computational complexity, and improves the generalizability of the model. Qualitative and quantitative comparative experimental results show that comprehensive performances of the model can catch up with or even surpass state‐of‐the‐art (SOTA) algorithms in terms of processing speed, subjective visual perception and objective evaluation metrics. In addition, application test results prove that DGC‐UWnet can be used as the pre‐processing for underwater applications of other visual algorithms such as improving performance of YOLOv5l.

B2-Net: an artificial intelligence powered machine learning framework for the classification of pneumonia in chest x-ray images

A chest x-ray radiograph is still the global standard for diagnosing pneumonia and helps distinguish between bacterial and viral pneumonia. Despite several studies, radiologists and physicians still have trouble correctly diagnosing and classifying pneumonia without false negatives. Modern mathematical modeling and artificial intelligence could help to reduce false-negative rates and improve diagnostic accuracy. This research aims to create a novel and efficient multiclass machine learning framework for analyzing and classifying chest x-ray images on a graphics processing unit (GPU). Researchers initially applied a geometric augmentation using a positional transformation function to the original dataset to enhance the sample size and aid future transfer learning. Models with the best accuracy, area under the receiver operating characteristics (AUROC), F1 score, precision, recall, and specificity are chosen from a pool of nine state-of-the-art neural network models. The best-performing models are then retrained using an ensemble technique using depth-wise convolutions, demonstrating significant improvements over the baseline models employed in this research. With a remarkable 97.69% accuracy, 100% recall, and 0.9977 AUROC scores, the proposed Bek-Bas network (B2-Net) model can differentiate between normal, bacterial, and viral pneumonia in chest x-ray images. A superior model is retrained using the chosen dense convolutional network-160, residual network-121, and visual geometry group network-16 ensemble models. The diagnostic accuracy of the x-ray classification unit is enhanced by the newly designed multiclass network, the B2-Net model. The developed GPU-based framework has been examined and tested to the highest clinical standards. After extensive clinical testing, the final B2-Net model is implemented on an NVIDIA Jetson Nano GPU computer. Healthcare facilities have confirmed the B2-Net is the most effective framework for identifying bacterial and viral pneumonia in chest x-rays.

Acceleration Techniques for Automated Design of Approximate Convolutional Neural Networks

The main issue connected with using approximate components such as approximate multipliers in deep convolutional neural networks (CNN) during the design process is the necessity to emulate them due to the lack of native support for approximate operations in modern CPUs and GPUs, which is computationally expensive. To accelerate the emulation of approximate operations of CNNs on GPUs, we propose TFApprox4IL, a software library supporting both symmetric as well as asymmetric quantization modes, approximate 8xN bit multipliers emulated using lookup tables, a new type of approximate layer known as approximate depthwise convolution, and quantization-aware training. The TFApprox4IL performance is extensively evaluated in the simulation of approximate implementations of MobileNetV2 and ResNet networks on Nvidia Pascal and Tesla GPU architectures. Furthermore, TFApprox4IL is also evaluated in neural architecture search (NAS) algorithms to automatically design CNN architectures that directly employ approximate multipliers. On two different NAS methods, EvoApproxNAS and Google Model Search (GMS), we show how approximate multipliers can effectively be incorporated into the CNN design process. To estimate the energy consumption of the approximate CNNs, AxMultAT tool based on Timeloop and Accelergy is introduced. Contrasted to the highly optimized GPU-based CNN simulation implemented using exact arithmetic operations available within TensorFlow, the average overhead of the inference and training, introduced by TFApprox4IL, is <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$13.6\times $ </tex-math></inline-formula> and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$8.0\times $ </tex-math></inline-formula> , respectively, considering ResNet50V2 and MobileNetV2 CNNs on ImageNet and CIFAR-10 data sets. This overhead was reduced by one order of magnitude with respect to previous methods.

EADD-YOLO: An efficient and accurate disease detector for apple leaf using improved lightweight YOLOv5.

Current detection methods for apple leaf diseases still suffer some challenges, such as the high number of parameters, low detection speed and poor detection performance for small dense spots, which limit the practical applications in agriculture. Therefore, an efficient and accurate model for apple leaf disease detection based on YOLOv5 is proposed and named EADD-YOLO. In the EADD-YOLO, the lightweight shufflenet inverted residual module is utilized to reconstruct the backbone network, and an efficient feature learning module designed through depthwise convolution is proposed and introduced to the neck network. The aim is to reduce the number of parameters and floating point of operations (FLOPs) during feature extraction and feature fusion, thus increasing the operational efficiency of the network with less impact on detection performance. In addition, the coordinate attention module is embedded into the critical locations of the network to select the critical spot information and suppress useless information, which is to enhance the detection accuracy of diseases with various sizes from different scenes. Furthermore, the SIoU loss replaces CIoU loss as the bounding box regression loss function to improve the accuracy of prediction box localization. The experimental results indicate that the proposed method can achieve the detection performance of 95.5% on the mean average precision and a speed of 625 frames per second (FPS) on the apple leaf disease dataset (ALDD). Compared to the latest research method on the ALDD, the detection accuracy and speed of the proposed method were improved by 12.3% and 596 FPS, respectively. In addition, the parameter quantity and FLOPs of the proposed method were much less than other relevant popular algorithms. In summary, the proposed method not only has a satisfactory detection effect, but also has fewer parameters and high calculation efficiency compared with the existing approaches. Therefore, the proposed method provides a high-performance solution for the early diagnosis of apple leaf disease and can be applied in agricultural robots. The code repository is open-sourced at https://github.com/AWANWY/EADD-YOLO.

A Hybrid Feature Extraction and Classification using Xception-RF for Multiclass Disease Classification in Plant Leaves

ABSTRACT In today’s world, agriculture is one of the most important assets as it affects the economy of the country. Biotic stress caused by fungi, bacteria, and viruses is common in plants; hence plant disease recognition is essential as it could identify diseases at the early stages of infection. Most of the previous plant disease works involve PlantVillage dataset, which is taken in a controlled environment. In this paper, the Banana leaf dataset is used which is combined from two datasets. CNNs are becoming more popular in many forms, such as pre-trained models, and in this research, we employed a pre-trained model in conjunction with an ML classifier to predict plant disease. Pre-trained models such as VGG 16, RESNET, MobileNet, InceptionResNetV2, Inception, and Xception models were deployed for feature extraction. Based on the above models, the proposed Xception-RF model achieved 100% accuracy with no misclassified classes. Using an Xception model with depth-wise convolutions reduces the number of parameters and the cost of implementation. The proposed work merges machine learning and deep learning by extracting features using CNN and then using ML-based RF for classification, which has resulted in an increase in performance in terms of accuracy.

A Machine Learning Approach for the Estimation of Severity of Psoriasis Disorder Using Depth-Wise Convolution Neural Network

Objectives: To design and develop a computer-aided tool for the estimation of the severity of psoriasis-affected skin areas using depth-wise convolution neural networks. Methods: The experiment consists of 5951 input images which are divided into a training set, validation set, and testing set. The training set consists of 2282 mild images, 646 moderately affected images, and 323 severely affected psoriasis images. The validation set consists of 600 mild, moderate, and severely affected images in each category. The testing set consists of 200 mild, moderate, and severely affected images in each category. The reconstructed MobileNet machine learning model is used in the estimation of severity. The reason for using the reconstructed MobileNet machine learning model is the less turnaround time required in the estimation of severity and the use of depth-wise convolution layers around which the model is built, which can be implemented in low-power, low-memory handheld devices. At the initial stages, pre-trained weights of ImageNet are used and at a later stage of the experiment, the network learns from our input data set. Findings: The experimental results proved to be significant and accurate when verified with the dermatologists. The fully trained network on the input data set is found to predict the severity around 90% which is counter-verified by the dermatologists. At the beginning of the experiments, when the neural network was run on the data set with the weights from ImageNet, the estimate of the disorder was not convincing. As the neural network was trained and weights were recalculated based on our collected data set, the results got improved. Even though MobileNet machine learning model was used, the average turnaround time for 20 epochs was about 27 minutes, which is a significant part of the experiment in developing the tool. An average F1 score of .94 is achieved when all the layers of the network are trained. Novelty: The novelty of the work lies in the prediction of the types of the severity psoriasis disorder accurately with less turnaround time with an accuracy of 90% where the results are cross verified by dermatologists and the tool canbe implemented on low-powered, low-processing hand-held devices. Keywords: Psoriasis; Depthwise convolution neural network; Machine Learning; Keras; Skin Texture

FBMSNet: A Filter-Bank Multi-Scale Convolutional Neural Network for EEG-Based Motor Imagery Decoding.

Motor imagery (MI) is a mental process widely utilized as the experimental paradigm for brain-computer interfaces (BCIs) across a broad range of basic science and clinical studies. However, decoding intentions from MI remains challenging due to the inherent complexity of brain patterns relative to the small sample size available for machine learning. This paper proposes an end-to-end Filter-Bank Multiscale Convolutional Neural Network (FBMSNet) for MI classification. A filter bank is first employed to derive a multiview spectral representation of the EEG data. Mixed depthwise convolution is then applied to extract temporal features at multiple scales, followed by spatial filtering to mitigate volume conduction. Finally, with the joint supervision of cross-entropy and center loss, FBMSNet obtains features that maximize interclass dispersion and intraclass compactness. We compare FBMSNet with several state-of-the-art EEG decoding methods on two MI datasets: the BCI Competition IV 2a dataset and the OpenBMI dataset. FBMSNet significantly outperforms the benchmark methods by achieving 79.17% and 70.05% for four-class and two-class hold-out classification accuracy, respectively. These results demonstrate the efficacy of FBMSNet in improving EEG decoding performance toward more robust BCI applications. The FBMSNet source code is available at https://github.com/Want2Vanish/FBMSNet.

ForensicNet: Modern convolutional neural network-based image forgery detection network.

The advancements in Image editing techniques produce realistic-looking artificial images with ease. These images can easily circumvent the forensic systems making the authentication process more tedious and difficult. To overcome this problem, we introduce a modern convolutional neural network (CNN) named ForensicNet, inspired by the recent developments in computer vision. The three major contributions of our CNNs are inverted bottleneck, separate downsampling layers, and using depth-wise convolutions for mixing information in the spatial dimension. The inverted bottlenecks help improve accuracy and reduce network parameters/FLOPs. The separate downsampling layers help converge the network. The normalization layers also help stabilize training whenever the spatial resolution is changed. The depth-wise convolution is a grouped convolution where the number of groups and channels are the same. The experiments show that ForensicNet outperforms the state-of-the-art methods by a large margin.

Lightweight object detection algorithm based on YOLOv5 for unmanned surface vehicles

Visual detection technology is essential for an unmanned surface vehicle (USV) to perceive the surrounding environment; it can determine the spatial position and category of the object, which provides important environmental information for path planning and collision prevention of the USV. During a close-in reconnaissance mission, it is necessary for a USV to swiftly navigate in a complex maritime environment. Therefore, an object detection algorithm used in USVs should have high detection s peed and accuracy. In this paper, a YOLOv5 lightweight object detection algorithm using a Ghost module and Transformer is proposed for USVs. Firstly, in the backbone network, the original convolution operation in YOLOv5 is upgraded by convolution stacking with depth-wise convolution in the Ghost module. Secondly, to exalt feature extraction without deepening the network depth, we propose integrating the Transformer at the end of the backbone network and Feature Pyramid Network structure in the YOLOv5, which can improve the ability of feature expression. Lastly, the proposed algorithm and six other deep learning algorithms were tested on ship datasets. The results show that the average accuracy of the proposed algorithm is higher than that of the other six algorithms. In particular, in comparison with the original YOLOv5 model, the model size of the proposed algorithm is reduced to 12.24 M, the frames per second reached 138, the detection accuracy was improved by 1.3%, and the mean of average precision (0.5) reached 96.6% (from 95.3%). In the verification experiment, the proposed algorithm was tested on the ship video collected by the “JiuHang 750” USV under different marine environments. The test results show that the proposed algorithm has a significantly improved detection accuracy compared with other lightweight detection algorithms.

Computationally Lightweight Hyperspectral Image Classification Using a Multiscale Depthwise Convolutional Network With Channel Attention

Convolutional networks have been widely used for the classification of hyperspectral images; however, such networks are notorious for their large number of trainable parameters and high computational complexity. Additionally, traditional convolution-based methods are typically implemented as a simple cascade of a number of convolutions using a single-scale convolution kernel. In contrast, a lightweight multiscale convolutional network is proposed, capitalizing on feature extraction at multiple scales in parallel branches followed by feature fusion. In this approach, 2D depthwise convolution is used instead of conventional convolution in order to reduce network complexity without sacrificing classification accuracy. Furthermore, multiscale channel attention is also employed to selectively exploit discriminative capability across various channels. To do so, multiple 1D convolutions with varying kernel sizes provide channel attention at multiple scales, again with the goal of minimizing network complexity. Experimental results reveal that the proposed network not only outperforms other competing lightweight classifiers in terms of classification accuracy but also exhibits a lower number of parameters as well as significantly less computational cost.

Faster Metallic Surface Defect Detection Using Deep Learning with燙hannel燬huffling

Deep learning has been constantly improving in recent years, and a significant number of researchers have devoted themselves to the research of defect detection algorithms. Detection and recognition of small and complex targets is still a problem that needs to be solved. The authors of this research would like to present an improved defect detection model for detecting small and complex defect targets in steel surfaces. During steel strip production, mechanical forces and environmental factors cause surface defects of the steel strip. Therefore, the detection of such defects is key to the production of high-quality products. Moreover, surface defects of the steel strip cause great economic losses to the high-tech industry. So far, few studies have explored methods of identifying the defects, and most of the currently available algorithms are not sufficiently effective. Therefore, this study presents an improved real-time metallic surface defect detection model based on You Only Look Once (YOLOv5) specially designed for small networks. For the smaller features of the target, the conventional part is replaced with a depth-wise convolution and channel shuffle mechanism. Then assigning weights to Feature Pyramid Networks (FPN) output features and fusing them, increases feature propagation and the network’s characterization ability. The experimental results reveal that the improved proposed model outperforms other comparable models in terms of accuracy and detection time. The precision of the proposed model achieved by mAP@0.5 is 77.5% on the Northeastern University, Dataset (NEU-DET) and 70.18% on the GC10-DET datasets.

Automating mosquito taxonomy by compressing and enhancing a feature fused EfficientNet with knowledge distillation and a novel residual skip block

Identifying lethal vector and non-vector mosquitoes can become difficult for a layperson and sometimes even for experts, considering their visual similarities. Recently, deep learning (DL) became a solution to assist in differentiating the two mosquito types to reduce infections and enhance actions against them. However, the existing methods employed to develop a DL model for such a task tend to require massive amounts of computing resources and steps, making them impractical. Based on existing methods, most researchers rely on training pre-trained state-of-the-art (SOTA) deep convolutional neural networks (DCNN), which usually require about a million parameters to train. Hence, this method proposes an approach to craft a model with a far lower computing cost while attaining similar or even significantly better performance than pre-existing models in automating the taxonomy of several mosquitoes. This method combines the approach of layer-wise compression and feature fusion with enhanced residual learning that consists of a self-normalizing activation and depthwise convolutions.•The proposed method yielded a model that outperformed the most recent and classic state-of-the-art deep convolutional neural network models.•With the help of the modified residual block and knowledge distillation, the proposed method significantly reduced a fused model's cost while maintaining competitive performance.•Unlike other methods, the proposed method had the best performance-to-cost ratio.

Lightweight Method for Plant Disease Identification Using Deep Learning

In the deep learning approach for identifying plant diseases, the high complexity of the network model, the large number of parameters, and great computational effort make it challenging to deploy the model on terminal devices with limited computational resources. In this study, a lightweight method for plant diseases identification that is an improved version of the ShuffleNetV2 model is proposed. In the proposed model, the depthwise convolution in the basic module of ShuffleNetV2 is replaced with mixed depthwise convolution to capture crop pest images with different resolutions; the efficient channel attention module is added into the ShuffleNetV2 model network structure to enhance the channel features; and the ReLU activation function is replaced with the ReLU6 activation function to prevent the generation of large gradients. Experiments are conducted on the public dataset PlantVillage. The results show that the proposed model achieves an accuracy of 99.43%, which is an improvement of 0.6 percentage points compared to the ShuffleNetV2 model. Compared to lightweight network models, such as MobileNetV2, MobileNetV3, EfficientNet, and EfficientNetV2, and classical convolutional neural network models, such as ResNet34, ResNet50, and ResNet101, the proposed model has fewer parameters and higher recognition accuracy, which provides guidance for deploying crop pest identification methods on resource-constrained devices, including mobile terminals.