关键词: 图像处理, 病虫害, 图像识别, 卷积神经网络, 模型剪枝

Abstract:  The design of lightweight plant leaf pests and diseases identification algorithm is the key to identifying plant leaf pests and diseases on the mobile terminal. This research proposes a lightweight plant leaf pests and diseases identification algorithm SimplifyResNet based on the improved ResNet model. Using artificially collected images and PlantVillage dataset images as experimental data, the ResNet model is improved according to the actual requirements for accuracy, speed, and model size of plant pests and diseases identification on the mobile terminal. The design uses 5×5 convolution instead of 7×7 convolution, uses the bottleneck structure of the residual block to replace the shortcut structure, and uses model pruning to process the trained model. The SimplifyResNet model was tested with 5 786 images on the test set. It was proved that the bottleneck structure of 5×5 convolution and residual block can effectively reduce the number of model parameters, and model pruning can effectively reduce the model size after training. The SimplifyResNet model has an accuracy of 92.45% for image recognition in the test set, a recognition time of 48 ms, and a memory size of 36.14 Mb. Compared to models, such as LeNet, AlexNet, and MobileNet, the accuracy of this method is 18.3%, 7.45%, and 1.2% higher, respectively. This research solves the most critical algorithm design problem for mobile plant pests and diseases identification and makes functional explorations for mobile plant pests and diseases identification.

Key words:  image processing, disease, image recognition, convolution neural network, model pruning