Tobacco-disease Image Recognition via Multiple-Attention Classification Network

Abstract

The recognition of disease images is of great importance in modern tobacco agriculture. Compared with other plant diseases, the early disease spots of tobacco are highly similar, and the quantity of tobacco disease images is scarce due to the difficulty of data collection. As such, the image recognition of early tobacco diseases is challenging. In response to these problems, this study proposes a novel image-classification method, namely, Multiple-Attention Classification Network (MACN). It improves the existing image classification network from two aspects: 1) the acquisition of low-level features based on transfer-learning and 2) the extraction of high-level features by modelling the multiple dependencies in low-level features. The pretrained model is used to obtain the initial feature map, thereby breaking the limitation of small-scale dataset. A novel multiple-attention module (MAM) is designed to learn fine-grained differences between classes. To evaluate the proposed method, this paper conducts experiments on field-collected images. Results show that this method is superior in identifying early tobacco diseases and has demonstrated better robustness when dealing with the actual field images with complex factors.