Crop Disease Recognition Based on Modified Light-Weight CNN With Attention Mechanism

Abstract

The agricultural production is greatly affected by various plant diseases. Classifying the severity of crop diseases is the requirement for formulating disease prevention and control strategies. However, the differences between different severity of the same crop disease are very tiny. It increases the difficulty of correct crop disease recognition. For example, at the early stage of the disease, the lesions on the leaves are not obvious. And it is very difficult to extract the features of the lesions. However, these very small color and texture differences of the lesions are the key patterns to distinguish different kinds of diseases of the same species. In order to achieve better performance in the fine-grained classification of the crop diseases, a modified light-weight convolution neural network was proposed. Multi-scale convolution kernel and coordinate attention mechanism are introduced in SqueezeNext to extract the features of the lesions accurately. The performance of the proposed model was evaluated using the AI challenger 2018 plant disease recognition dataset, and the recognition accuracy can reach 91.94%, which is 3.02% point higher than the original SqueezeNext model. In order to verify the effectiveness of the proposed model, comparative experiments were carried out using ReseNet50, Xception and mobilenetv2. The experimental results showed that the accuracy of the proposed method was slightly better than Xception, while the model size is only 2.83 MB, which is only 3.45% of Xception. The proposed method balances the performance and efficiency very well. Thus, it is suitable for deployment on mobile terminals and other embedded resource-constrained devices, which help to promote the popularization of smart agriculture application.