基于改进YOLOX的棉花姿态品级识别及其定位研究  

Research on posture and grade recognition of cotton and its localization using improved YOLOX

作　　者：谢嘉[1] 陈学飞 李永国[1] 金昌兵 梁锦涛 孙帅浩 Xie Jia;Chen Xuefei;Li Yongguo;Jin Changbing;Liang Jintao;Sun Shuaihao(College of Engineering Science and Technology,Shanghai Ocean University,Shanghai 201306,China;Xiaojia Intelligent Technology(Shanghai)Company,Shanghai 201306,China;School of Mechano-Electronic Engineering,Xidian University,Xi’an 100083,China)

机构地区：[1]上海海洋大学工程学院,上海201306 [2]小家智能科技(上海)有限公司,上海201306 [3]西安电子科技大学机电工程学院,西安710071

出　　处：《棉花学报》2024年第4期328-339,共12页Cotton Science

基　　金：国家自然科学基金(51876114);国家自然科学基金青年项目(51605363)。

摘　　要：【目的】旨在解决高质量采棉要求下,采棉机对不同姿态和品级的棉花进行精确识别与定位的问题,提出了1种基于改进YOLOX的棉花检测方法YOLOX-Cotton。【方法】YOLOX-Cotton使用YOLOX模型作为主体框架,包含识别模块和定位模块,并引入了CA(coordinate attention)模块和SIoU损失函数,以多种姿态、品级的棉花图片作为数据集,对其进行训练并测试。【结果】YOLOX-Cotton模型的识别模块能够识别不同姿态和品级的棉花,且模型精确率、召回率和平均精度均值达到92.9%、86.8%和92.4%,与原YOLOX模型相比分别提升了5.2、5.5和6.1百分点。该模型的定位模块能够准确获得棉花的位置,测量结果均在田间试验验证结果的阈值范围内,所有样本的标准偏差均小于0.01。【结论】YOLOX-Cotton能够有效解决采棉机在高质量采棉要求下对棉花的识别与定位问题,将为实现高质量采棉提供了有力的技术支撑。[Objective]This paper aims to solve the problem of accurate recognition and localization of cotton with different postures and grades by cotton picker under the requirement of high-quality cotton picking.A cotton detection method YOLOX-Cotton based on the improved YOLOX is proposed.[Methods]YOLOX-Cotton uses YOLOX as the main framework,including a recognition module and a localization module,and incorporates coordinate attention(CA)module and SIoU loss function,and takes various posture and grade cotton pictures as data sets to train and test.[Results]The detection module of YOLOX-Cotton was capable of detecting cotton with different postures and grades,and the model precision,recall and average precision reached 92.9%,86.8%and 92.4%,which were improved by 5.2,5.5 and 6.1 percentage points,compared with the original YOLOX,respectively.The localization module of this model was capable of accurately obtaining the location of the cotton,the measurements were kept within the threshold range of the validated results of the field trial,and the standard deviation of all samples was less than 0.01.[Conclusion]The experiment proves that the YOLOX-Cotton can effectively solve the problem of cotton detection and localization by cotton picker under the requirement of high-quality cotton picking,and provides strong technical support for the realization of high-quality cotton picking.

关 键 词：棉花 目标检测 三维定位 注意力机制 损失函数 

分 类 号：S562[农业科学—作物学]