MACAT Net a hierarchical CNN transformer framework with multi scale attention for robust cassava leaf disease detection

Cassava is an important food crop to millions of people across the globe but its productivity is highly compromised by the leaf diseases like the Cassava Mosaic Disease, Cassava Brown Streak Disease and bacterial blight whose symptoms are subtle and have overlapping appearances when not controlled through imaging. The inconsistencies of illumination, complicated backgrounds, occlusions, and inter-class analogy are some of the primary factors that reduce the accuracy of standard image-based diagnosis systems. In order to overcome these limitations, the present paper introduces MACAT-Net, a new Multi-scale Attention-guided CNN-Transformer Network that can effectively and reliably detect multiple instances of cassava leaf disease using realistic samples of varied images. The presented framework combines the convolutional neural network-based backbone of hierarchical local feature extraction with the transformer-based global context modeling block that helps to identify the long-range spatial dependencies between leaf areas. A multi-scale attention system is proposed to discriminate diseases related patterns selectively at various receptive fields to allow fine grained discrimination between disease similarities at the image level and block-out background noise. Compared to the current one-representation methods, MACAT-Net is adaptively fusing convolutional and transformer representations, making sure that it learns complementary features on the texture-level, as well as global pattern distribution of symptoms. Extensive experiments are done on benchmark datasets of cassava leaf disease with realistic variability to determine the accuracy of classification, precision, recall, F1-score, and resistance to adverse visual conditions. The proposed MACAT-Net is invariably better than the latest state-of-the-art convolutional, transformer-only, and hybrid deep learning models, as it has better generalization and stability. The contribution of the multi-scale attention and hybrid fusion strategy is also further confirmed by the ablation studies. The findings reveal that MACAT-Net is a viable and scalable diagnostic system of automated cassava leaf disease that has a high potential in practical decision-support systems in precision agriculture.