文献类型： 会议论文

第一作者： Zekai Cheng

作者： Zekai Cheng ¹ ; Rongqing Huang ¹ ; Rong Qian ² ; Shu Zhang ¹ ; Xinzhuo Jiang ¹ ; Meifang Liu ¹ ; Xiwen Qu ¹ ;

作者机构： 1.School of Computer Science and Technology, Anhui University of Technology, Ma’anshan, China

2.Institute of Agricultural Economy and Information, Anhui Academy of Agricultural Sciences, Hefei, China

关键词： Degradation;Crops;Optimization methods;Big Data;Feature extraction;Internet of Things;Usability

会议名称： International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering

主办单位：

页码： 154-158

摘要： The practical application of crop pest detection methods has been limited by the large number of parameters and computations, and we built a lightweight crop pest detection method YOLOLite-CSG in our previous research, which basically removed this limitation. However, further analysis shows that YOLOLite-CSG still has problems that affect the performance in terms of the prior box generation method, downsampling method and bounding box regression loss function. In response to these problems, this paper proposes the prior box generation method based on label box oversampling and clustering (BS-Medians), the downsampling method based on parallel feature transformation (PSPD-Conv), and the bounding box regression loss function with size difference optimization capability (FIoU Loss). We optimize YOLOLite-CSG using the above methods to build YOLOLite-X, an optimized lightweight crop pest detection method. The experiment results show that YOLOLite-X exceeds YOLOLite-CSG in crop pest detection precision and is higher than other state-of-the-art methods. Meanwhile, YOLOLite- X has a smaller number of parameters and computations, which is more conducive to practical deployment and application.

分类号： tp3-53