Abstract: Traditional methods for detecting cotton pests and diseases rely on plant protection experts or experienced farmers, which is time-consuming and laborious, and has low accuracy in identifying small lesions that are difficult to observe by the human eye．This poses new challenges on how to use deep learning methods to improve recognition accuracy and efficiency, reduce labor costs, and accurately identify small lesions．To address this issue, a cotton pest detection algorithm based on improved YOLOv8 with high accuracy and good real-time performance was proposed．Firstly, DCNv3 structure was used to replace ordinary convolution in Bottleneck structure of YOLOv8 C2 module, forming a new module named C2f-DCNv3．Secondly, Gaussian Contextual Attention was added after the last C2f-DCNv3 module in the Head to improve accuracy of model while maintaining real-time detection capability．Experimental results on the open source CottonInspect cotton field insect recognition research image dataset showed that, the mAP of improved method was 0.706, the Inference Time was 0.6 ms, and the model size was 5.7 MB．Compared with original model YOLOv8n, the mAP was increased by 3.0, the Inference Time was increased by 0.1 ms, and model size remained basically unchanged, achieving accurate and efficient identification of cotton pests．