AUTHOR=Liu Mingxin , Liang Haofeng , Hou Mingxin 

TITLE=Research on cassava disease classification using the multi-scale fusion model based on EfficientNet and attention mechanism

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 13 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2022.1088531

DOI=10.3389/fpls.2022.1088531

ISSN=1664-462X

ABSTRACT=Cassava disease is one of the leading causes to the serious decline of cassava yield. Aimed at solving problems of different colors, abnormal shapes, and areas of cassava leaf disease spots, the authors of this article have studied the classification of cassava leaf disease by a deep convolutional neural network (CNN) and realized the recognition and classification of cassava disease by image classification technology. First, in the case of unbalanced disease data, focal loss (FL) was used to reduce the weight penalty of easily classified samples and strengthen the weight reward of difficult-to-classify samples, thus reducing the emphasis on most categories caused by unbalanced sample data. Second, the disease features of the cassava leaves were extracted using the Convolutional Block Attention Module (CBAM) to strengthen the feature classification of the disease, and the obtained feature map contains the key recognition information of the disease. Finally, an improved method based on multi-scale feature fusion was proposed to extract the characteristics of cassava disease by combining receptive fields of different sizes, which could not only obtain the information on cassava disease spots and colors but also obtain rich semantic information. The results showed that compared with that of the original model, the average recognition rate of the cassava leaf disease classification model based on multi-scale fusion was improved by nearly 4% and reached 88.1%, providing theoretical support and practical tools for the identification and early diagnosis of the diseased plant leaves.