Authors

Abstract

To address the low recognition accuracy and slow detection speed of cotton leaf pests and diseases in natural environments, a detection method based on an improved YOLOv5s model was proposed. The enhanced model integrates the Ghost module and the C3Faster module to increase inference speed and reduce model complexity, achieving lightweight performance without significantly compromising accuracy. To counteract the tendency of common cotton pest and disease features to be lost in complex natural scenes, a Coordinate Attention (CA) mechanism was introduced to improve the network's recognition and localization capabilities. The parameters, FLOPs, and weight file size of the improved model were reduced to 65.5%, 66.2%, and 67.1% of those of the original YOLOv5s model, respectively. On a self-built dataset, the improved YOLOv5s model achieved a mean average precision (mAP) improvement of 10.5%, 0.2%, and 0.4% compared to YOLOv4, YOLOv5s, and YOLOv7, respectively. The model was deployed on a Jetson Orin NX development board with CUDA acceleration, achieving a real-time detection speed of 76.3 frames per second.

Abstract in Chinese

针对自然环境中棉花叶片病虫害识别精度低、识别速度慢的问题，提出了一种基于改进YOLOv5s的检测方法。该方法基于YOLOv5s模型，通过结合Ghost模块和C3Faster模块来提高推理速度，实现模型轻量化，而不会显著影响准确性。由于棉花常见病虫害的特征在自然环境中很容易丢失，因此添加了CA注意机制模块，以帮助轻量级网络提高识别和定位能力。改进后的模型参数、FLOP和权重文件大小分别为YOLOv5s的65.5%、66.2%和67.1%。将自建数据集与YOLOv4、YOLOv5s和YOLOv7进行比较，结果表明，改进的YOLOv5模型的平均准确率分别提高了10.5%、0.2%和0.4%。改进后的模型部署在Jetson Orin NX开发板上，并使用CUDA加速。加速模型检测速度为76.3帧/秒。