Abstract: With the popularity and rapid development of deep learning applications, image recognition methods based on deep learning are widely used in the field of crop diseases and insect pests. However, most neural networks attach importance to the improvement of recognition accuracy, but ignore the huge parameter computation amount of neural networks. In order to solve this problem, based on the progressive growing of GANs discriminator model and convolutional attention module. an improved CPDM network model was proposed to identify crop pests and diseases. By adjusting the network structure of the progressive growing of GANs discriminator, the feature extraction capability of CPDM network model was enhanced by using balanced learning rate, pixellevel feature vector normalization and convolutional attention module, and the recognition accuracy of real images was improved. The experiment was carried out on the PlantVillage dataset, and compared with VGG16, VGG19 and ResNet18, the TOP1 accuracy was 99.06%, 96.50%, 96.65% and 98.86%, respectively, which was improved by 2.56%, 2.41% and 0.2%, respectively. And the number of parameters was only 8.2 M. The experimental results show that the proposed CPDM network model meets the purpose of effectively controlling the calculation amount of neural network parameters on the basis of ensuring the classification accuracy.

Key words: crop diseases and pests, progressive growing of GANs, CBAM, finegrained classification

摘要： 随着深度学习应用的普及和飞速发展，基于深度学习的图像识别方法广泛应用于农作物病虫害领域，但大部分的神经网络重视识别准确率的提高，却忽略神经网络庞大的参数计算量。为解决这个问题，基于渐进式生成对抗网络判别器模型和卷积注意力模块，提出一种改进的渐进式生成对抗网络判别器CPDM网络模型对农作物病虫害进行识别。通过对渐进式生成对抗网络判别器网络结构的调整，采用均衡学习率、像素级特征向量归一化和卷积注意力模块增强CPDM网络模型的特征提取能力，提高对真实图片的识别准确率。试验在PlantVillage数据集上进行，将该模型与VGG16、VGG19和ResNet18进行比较，得到TOP1准确率分别为99.06%、96.50%、96.65%、98.86%，分别提高2.56%、2.41%、0.2%，且参数量仅为8.2 M。试验证明提出的CPDM网络模型满足在保证分类准确率的基础上，有效控制神经网络参数计算量的目的。

关键词: 农作物病虫害, 渐进式生成对抗网络, 卷积注意力模块, 细粒度分类