弱光条件下猪舍清洗目标检测  

Detection of piggery cleaning target under low-light condition

作　　者：李颀[1] 李煜哲 LI Qi;LI Yuzhe(College of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi an 710021,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,陕西西安710021

出　　处：《浙江农业学报》2023年第9期2240-2249,共10页Acta Agriculturae Zhejiangensis

基　　金：陕西省科技厅项目(S2023-YF-YBNY-0232)。

摘　　要：无人工干预的猪舍清洗可以有效预防猪只感染疾病。针对猪舍场景中光线较弱、特征不明显、存在相互遮挡情况导致对清洗目标的检测准确率低的问题,提出一种改进的YOLOv5算法。在预处理阶段,利用基于双边滤波的Retinex算法提高在弱光条件下对猪舍清洗目标的检测能力;在清洗目标检测过程中,通过在YOLOv5网络的backbone中引入CBAM(convolutional block attention module)注意力机制,使网络学习到猪舍清洗目标更多的有效特征信息;对网络中边框回归损失函数进行改进,采用DIoU-NMS算法筛选出猪舍清洗目标的候选框,提高在部分遮挡情况下清洗目标的检测精度。实验结果表明:在猪舍清洗目标的测试集上改进后的YOLOv5目标检测算法较基准算法准确率提高7.3百分点,召回率提高7.6百分点,平均准确度提高7.1百分点,在弱光条件和目标遮挡情况下鲁棒性更高。研究结果为畜牧智能清洗设备的研发提供了基础。The cleaning of piggery without manual intervention can effectively prevent pigs from contracting diseases.Aiming at the problems of weak light,inconspicuous features,and mutual occlusion in the piggery scene,the low detection accuracy of cleaning targets,an improved YOLOv5 algorithm was proposed.In the preconditioning stage,the Retinex algorithm based on bilateral filtering was used to improve the detection ability of the piggery cleaning target under low light conditions;In the cleaning target detection process,the CBAM(convolutional block attention module)attention mechanism was introduced into the backbone of the YOLOv5 network to enable the network to learn more effective feature information of the cleaning target in the piggery;The frame regression loss function in the network was improved,and the DIoU-NMS algorithm was used to filter out the candidate frame of the cleaning target in the piggery,which improved the detection accuracy of the cleaning target in the case of partial occlusion.The experimental results showed that the improved YOLOv5 target detection algorithm improved the accuracy rate by 7.3 percentage point,the recall rate by 7.6 percentage point,and the average accuracy by 7.1 percentage point compared with the benchmark algorithm on the test set of piggery cleaning targets.The robustness was higher in the case of target occlusion.The result provided a basis for the research and development of animal husbandry intelligent cleaning equipment.

关 键 词：猪舍 清洗目标 弱光 注意力机制 YOLOv5算法 

分 类 号：S237[农业科学—农业机械化工程]