A deep feature fusion network using residual channel shuffled attention for cassava leaf disease detection

Cassava is a significant source of carbohydrates for tropical populations. However, diseases caused by agents such as bacteria, viruses, fungi, and phytoplasmas cause considerable economic damage to these crops. Existing methods for cassava disease detection require farmers to seek the assistance of agricultural experts for visual inspection and diagnosis, which is challenging and laborious. Most studies have employed pre-trained convolutional neural networks to detect diseases in cassava leaves. Also, it is essential to design customized deep neural networks specific to the target domain for precise classification. This research proposes a novel deep fusion of two networks, residual channel shuffled attention network and Efficientnet. The first network, RCSANet, was presented to capture contextual information using depthwise separable convolution effectively. It also integrates significant inter-spatial and inter-channel information using the triplet attention module and employs shuffled group convolution to capture features from distinct filter groups. As a result of incorporating the above architectural enhancements, the proposed feature fusion network exhibited better performance than the existing studies. The proposed network was trained on the Kaggle cassava leaf disease dataset with 21,367 samples and yielded a classification accuracy of 93.25%.