High-precision cotton pest detection algorithm based on improved YOLOv9c

To address the challenges of identifying cotton pests and diseases, especially due to their small size, similar appearance, and overlapping symptoms, this study introduces the SPC-YOLO object detection algorithm based on the YOLOv9c model. Key issues with YOLOv9c include its high computational load and limited small object detection capability. The proposed approach involves augmenting the dataset with techniques like Mixup, brightness enhancement, and translation to create a comprehensive dataset. The YOLOv9c backbone was replaced with StarNet for reduced computational complexity and improved small object detection. PConv from fasterNet was integrated into the head to enhance feature capture, and CARAFE upsampling replaced the original upsampling operator to improve reconstruction quality. Experimental results showed that SPC-YOLO achieved a detection accuracy of 91.0%, a recall rate of 84.5%, and a mean average precision (mAP) of 88.3%, with improvements of 1.8%, 4.8%, and 1.9% over the original YOLOv9c, respectively. The model also reduced parameters by 15.9% and floating-point operations (FLOPs) by 21.8%. SPC-YOLO outperforms networks such as YOLOv7, YOLOv8, YOLOv10, and RT-DETR, providing an effective solution for accurate cotton pest and disease detection.