Improved YOLO11 for Vision-Based Pest and Disease Detection in Chili Peppers

To address the low detection accuracy of small objects in complex backgrounds, this study proposes an improved model for pepper crop pest and disease detection, termed RSA-YOLO. Built upon YOLO11, the proposed model incorporates a super convolutional block attention module (SPCBAM) and a reparameterized heterogeneous multiscale structure (RepHMS) to enhance detection accuracy and robustness. Specifically, RepHMS integrates global heterogeneity modeling and flexible kernel selection, improving feature representation and adaptability to diverse target characteristics. Meanwhile, SPCBAM replaces the original C2PSA structure, enabling the network to focus on salient regions while suppressing background interference. Experimental results show that RSA-YOLO achieves an average precision of 80.5% on the test set, representing an 8% improvement over YOLO11, along with notable gains in recall and F1 score. In addition, the proposed model reduces the number of parameters by approximately 0.1 million, improving computational efficiency and maintaining a lightweight architecture. Overall, RSA-YOLO demonstrates strong detection performance and practical potential for small-object pest and disease detection in agricultural scenarios.