Attention-Based Approach for Cassava Leaf Disease Classification in Agriculture

Abstract

Cassava is a food crop that is rich in carbohydrates. However, the crop is vulnerable to many diseases. Research has revealed that image recognition using machine learning and deep learning techniques can be applied in automatic identification of cassava leaf diseases. Therefore this study focuses on using strongly discriminative features of the leaf regions affected by disease and weakening regions of low interest to improve the classification accuracy. A convolutional block attention module (CBAM) is a common attention mechanism integrated in feed-forward convolutional neural networks. In this study, CBAM is added to the pretrained ResNet50 and VGG19 models to recognize the cassava leaf regions affected by disease. This is done by sequentially inferring attention maps along two dimensions, channel and spatial for every intermediate feature map. The attention maps are then multiplied to the input feature map for adaptive feature refinement. The performance of baseline models such as EfficientNet, ResNet50, Inceptionv3, and Xception is compared with the attention-based models trained, validated and tested on a public dataset from Makerere University AI laboratory. ResNet50+CBAM achieve the highest performance with accuracy of 97%, precision of 96%, recall of 94% and F-measure of 95%. Conclusively, attention-based models perform better than the baseline models with a performance improvement of over 1%.