SMA-YOLO: An enhanced architecture for the detection of corn diseases based on YOLOv12

To address the issues of low effectiveness and accuracy, particularly with the detection of small objects in corn disease identification, this paper proposes enhancements to the You Only Look Once version 12 (YOLOv12) model to improve its detection capabilities for corn diseases. Initially, the StarNet module with a dynamic projection head is integrated into the backbone network, optimizing the YOLOv12 backbone. The resulting model is lightweight and well-suited for detection efficiency. Secondly, the A2C2f module of YOLOv12 incorporates a self-attention mechanism that combines cosine and Manhattan distances. This mechanism effectively captures the relationship between local and global features in the image, enhancing the detection accuracy for corn diseases, especially small object lesions. Additionally, the Adaptive Kernel Convolution (AKConv) block is introduced into the C3k2 module to enhance its feature extraction capability, effectiveness, and accuracy of detection. Experimental results demonstrate that the enhanced YOLOv12 model, which is integrated with StarNet, Manhattan self-attention module and AKConv block, namely SMA-YOLO, has achieved significant performance improvements in corn disease detection. The SMA-YOLO not only improves effectiveness and accuracy but also enhances identification in small object detection across various agricultural environments. Comparing the results from the original YOLOv12 to the enhanced YOLOv12 model, the detection time decreases from 9.4 ms/pic to 4.3 ms/pic, the recall rate increases from 73.88 % to 78.37 %, and the mean Average Precision ( mAP ) rises from 78.97 % to 83.07 %. These quantified results demonstrate its effectiveness and accuracy for precision agriculture.