Abstract: To address the problems of difficult small-object recognition, insufficient feature fusion, and complex background interference in plant disease detection, a multi-scale feedback aggregation detection model based on improved YOLOv12 was proposed. The model was validated on the PlantDoc dataset containing images of 30 categories of plant leaf diseases to improve the detection accuracy and robustness of lesion regions in complex environments. By introducing the Multi-Scale Enhanced Multi-Branch Aggregation with Feedback（MEMBA-F） module into the YOLOv12 framework, the improved model employed multi-branch convolution and a bidirectional information feedback mechanism to achieve adaptive cross-layer feature fusion, thereby enhancing the representation of lesion edges and fine-grained features. The results showed that the Precision, Recall, mAP50, and mAP50-95 of the improved model on the PlantDoc dataset reached 0.573, 0.372, 0.582, and 0.365, respectively, which were 7.7%, 1.4%, 3.7%, and 7.7% higher than those of YOLOv8, while maintaining an inference speed of 27.4 ms, demonstrating real-time detection capability. Compared with the baseline YOLOv12 model, the metrics of the improved model were further enhanced, verifying the effectiveness of the MEMBA-F module in multi-scale feature aggregation, cross-layer information interaction, and the representation of small objects and fine-grained features. The improved model balanced detection accuracy and efficiency, providing a reference for plant disease recognition in smart agriculture.

Key words: plant disease detection, YOLOv12, multi-scale feature fusion, MEMBA-F module