Identification of Maize Diseases Based on Dynamic Convolution and Tri-Attention Mechanism
Feilong Tang, Rosalyn R. Porle, Hoe Tung Yew, Farrah Wong
Abstract
Accurate, non-destructive classification of maize diseases is crucial for efficiently managing agricultural losses. While existing methods perform well in controlled environment dataset like PlantVillage, their accuracy often declines in real-world scenarios. In this work, ResNet50 is enhanced by integrating a dynamic convolution module and triplet attention modules. This method adaptively recalibrates the convolution kernel weights, establishing dependencies across spatial and channel dimensions through tensor rotation and residual transformations. The proposed method surpasses state-of-the-art alternatives, reaching 98.79% validation accuracy on the PlantVillage maize dataset and 97.47% on the Corn Leaf Disease Dataset through cross-validation. Even with complex backgrounds, it attains an average accuracy of 88.33% for classifying six types of maize diseases. Experimental results confirm its effectiveness in maize disease detection.