WTAD-YOLO: A lightweight tomato leaf disease detection model based on YOLO11

Abstract

Accurate localization of lesion regions is essential for the recognition of tomato leaf diseases. However, existing deep learning models face significant challenges in detecting small lesions in images, often resulting in reduced recognition accuracy. Meanwhile, their substantial computational resource consumption further restricts their practical deployment. This study proposes a novel tomato leaf disease detection model named WTAD-YOLO (Wavelet Transform ADown DySample YOLO) to address these limitations. Specifically, a C3k2_WTConv feature extraction module is designed to enhance multi-scale feature perception while only slightly increasing parameters. An ADown downsampling module is employed to reduce computational load and parameter count, while the DySample upsampling module ensures accurate multi-scale feature integration and efficient reconstruction of comprehensive information. Experimental results indicate that WTAD-YOLO consistently outperforms the baseline YOLO11 in detecting tomato leaf diseases, albeit with modest gains. The model attains a [email protected] of 0.917, an F1-score of 0.891, has 2.32M parameters, and a computational cost of 6.3 GFLOPs. In comparison to YOLO11, the [email protected] and F1-score exhibit enhancements of 1.9% and 2.0%, respectively, while the parameter count is diminished by about 10.0%. Meanwhile, GFLOPs remain unchanged. Furthermore, the model exhibits the least performance degradation in Domain Shift experiments. The proposed model outperforms common YOLO series models in detection performance, while maintaining relatively low computational and memory demands. Consequently, WTAD-YOLO offers a robust and efficient approach for the practical detection of tomato leaf diseases.

Topics & Concepts