Multi-prototype Co-saliency Model for Plant Disease Detection

Fangfang Liang, Ye Zhang, Xiaodan Zhang, Qing En

Abstract: Plant disease detection is crucial for mitigating significant reductions in crop yield. However, existing methods often misclassify healthy regions as diseased areas and struggle to detect small lesions, leading to the phenomenon of over-segmentation. To address these challenges, we propose a novel multi-prototype co-saliency model for plant disease detection. This model generates multiple differentiated prototypes and detects co-occurring disease regions in a co-saliency manner. It can learn more distinguishable disease representations in the embedding space and better excavate the implicit commonality information among different plant disease images. Specifically, we employ a response maps generation module that utilizes residual feature generation, seed selection, and democratic response mechanisms to produce response maps. This module enables the model to capture common attributes across different images effectively. Moreover, we design a multiple prototypes generation module that generates differentiated prototypes from multiple images using response maps and residual features. These prototypes are considered to be representative characteristics in the embedding space, guiding the model to focus on the diseased area effectively. Finally, we introduce a self-contrastive learning module that establishes contrastive relationships among prototypes in the embedding space. This module captures homogeneous and heterogeneous information, encouraging foreground prototypes to become similar to each other while promoting dissimilarity between foreground and background prototypes. Experimental results on the plant leaf disease and co-saliency datasets demonstrate that the proposed model achieves state-of-the-art performance.