Grape disease image classification based on lightweight convolution neural networks and channelwise attention

Abstract

In this paper, a lightweight convolution neural network model is proposed to diagnose grape diseases, including black rot, black measles and leaf blight. Focusing on small and low-latency models carried on mobile devices, a novel method is proposed based on lightweight convolution neural networks applying the channelwise attention (CA) mechanism. ShuffleNet V1 and V2 are chosen as the backbones, and squeeze-and-excitation (SE) blocks are considered as a CA mechanism to improve the ShuffleNet architecture. The proposed model is verified by an open dataset which includes 4,062 grape leaf images from four classes, including 3 diseased classes and 1 healthy class. The results of the experiments indicate the effectiveness of the proposed method. The best trained model accuracy is 99.14%, and the model size is compressed from 227.5 MB (AlexNet) to 4.2 MB.