DC-SE-DSCNet

Existing pest and disease identification models based on neural network image classification technology have high computational demands and prolonged processing times, which limit their widespread application. Therefore, further research on lightweight models is essential to improve recognition performance and enhance real-time processing capabilities. This study introduced a Dilated Convolution_Squeeze and Excitation_Depthwise Separable Convolution Network (DC-SE-DSCNet), which was a lightweight Convolutional Neural Network (CNN) designed for mobile devices and embedded systems in Mobile Edge Computing (MEC) environments. The model employed depthwise separable convolution (DSC) to replace traditional convolution, significantly reducing the parameter count and computational complexity. Additionally, Dilated Convolution (DC) expanded the receptive field, reduced the computational demands, and enhanced the feature extraction capabilities. The Squeeze-and-Excitation (SE) module was introduced as a channel attention mechanism to enhance the feature selection capabilities. This model integrated the max-pooling layer with the SE module to effectively preserve the critical feature information. The simulation results demonstrated that the DC-SE-DSCNet model achieved the accuracy of 94.14%, with a parameter count of only 0.17 M and the computational complexity of 0.48 GFLOPs. Overall, the DC-SE-DSCNet model effectively satisfied the requirements for low-cost, low-power, and high-performance devices.

现有基于神经网络图像分类技术的病虫害识别模型存在计算需求量大、处理耗时较长的问题，限制了其大规模推广应用。因此，开展轻量化模型的进一步研究，对于提升识别性能与实时处理能力至关重要。本研究提出了一种扩张卷积-压缩与激励-深度可分离卷积网络（Dilated Convolution_Squeeze and Excitation_Depthwise Separable Convolution Network，简称DC-SE-DSCNet），该网络是面向移动边缘计算（Mobile Edge Computing，MEC）环境下移动设备与嵌入式系统设计的轻量化卷积神经网络（Convolutional Neural Network，CNN）。该模型采用深度可分离卷积（Depthwise Separable Convolution，DSC）替代传统卷积，大幅降低了参数量与计算复杂度。此外，扩张卷积（Dilated Convolution，DC）可扩大感受野、降低计算需求并提升特征提取能力。本研究引入压缩与激励（Squeeze-and-Excitation，SE）模块作为通道注意力机制，以强化特征选择能力。该模型将最大池化层与SE模块相融合，可有效保留关键特征信息。仿真实验结果表明，DC-SE-DSCNet模型的识别准确率可达94.14%，参数量仅为0.17 M，计算复杂度仅为0.48 GFLOPs。综上，DC-SE-DSCNet模型可有效满足低成本、低功耗与高性能设备的应用需求。