Rice Disease Detection Method Based on CS-YOLOv8

To address the issue of insufficient detection accuracy caused by complex backgrounds, small lesion scales, and inconspicuous features in rice diseases, this paper proposes an improved lightweight object detection model, CS-YOLOv8, based on the YOLOv8n baseline model. Firstly, the CARAFE upsampling operator is introduced into the neck network, replacing the original upsampling module to enhance detail restoration capability for small-scale lesions. Secondly, the Spatial Group-wise Enhance (SGE) attention mechanism is introduced at the end of the backbone network to strengthen spatial saliency perception in diseased regions. Experimental results demonstrate that the improved CS-YOLOv8 model enhances <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{mAP}_{50}, \text{mAP}_{50: 95}$</tex>, recall rate, and precision by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$2.1 \%, 0.8 \%, 0.9 \%, 0.8 \%$</tex>, respectively, compared to the baseline model, while maintaining lightweight characteristics, thus meeting practical application requirements. Compared with YOLOv9n, Faster R-CNN and other models, CS-YOLOv8 has the accuracy increased by 1.3%-12.6%, the recall rate increased by 1.1 %-35.3 %, and the mean average precision increased by 0.9%-2.8%, which provides a more accurate solution for rice disease detection.