PAD-Net: a lightweight end-to-end detector for pepper anthracnose in complex agricultural environments.

Peppers are globally significant economic crops, but their yield is severely threatened by anthracnose. Automated detection remains challenging due to irregular lesion boundaries, drastic scale variations, complex backgrounds, and dense foliage occlusion. To address these issues, this study proposes PAD-Net, a lightweight end-to-end detector. Its LGMambaNet backbone synergizes DSwin-Mamba for deformable sliding window-based local modeling with FA-WTMamba for global frequency-domain analysis. A C3K2_PC module is introduced to capture multi-directional edge signals, enhancing sensitivity to blurred and multi-scale features. Furthermore, a DSSA module is designed to suppress background interference. Evaluated on the Pepper Anthracnose Dataset using standard COCO metrics, PAD-Net significantly outperforms the baseline DEIM model. Specifically, the PAD-Net nano version achieves an AP of 0.430, an AP 50 of 0.846, and an AP 75 of 0.382, representing a notable improvement over the baseline. The PAD-Net small version further attains an AP of 0.446, an AP 50 of 0.861, and an AP 75 of 0.404, delivering a substantial gain of 3.0% in AP compared to the baseline while also reducing model parameters. Generalization tests on Tomato Leaf, FieldPlant, and IP102 datasets further validate its superior detection capabilities and feature extraction. This research achieves an optimal balance between precision and efficiency, providing a reliable solution for automated monitoring and precision crop protection in smart agriculture.