基于改进YOLOv7和图像融合的哺乳期死亡仔猪检测方法  

The detection method of lactating dead pigs based on improved YOLOv7 and image fusion

作　　者：衡熙 沈明霞[2,4,5] 刘龙申 姚文[3,4,5] 李鹏 HENG XI;SHEN Mingxia;LIU Longshen;YAO Wen;LI Peng(College of Engineering,Nanjing Agricultural University,Nanjing 210031,China;College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210031,China;College of Animal Science&Technology,Nanjing Agricultural University,Nanjing 210095,China;Key Laboratory of Breeding Equipment,Ministry of Agriculture and Rural Affairs,Nanjing 210031,China;Jiangsu Smart Animal Husbandry Equipment Technology Innovation Center,Nanjing 210031,China)

机构地区：[1]南京农业大学工学院,江苏南京210031 [2]南京农业大学人工智能学院,江苏南京210031 [3]南京农业大学动物科技学院,江苏南京210095 [4]农业农村部养殖装备重点实验室,江苏南京210031 [5]江苏智慧牧业装备科技创新中心,江苏南京210031

出　　处：《南京农业大学学报》2025年第2期464-475,共12页Journal of Nanjing Agricultural University

基　　金：科技创新2030——“新一代人工智能”重大项目(2021ZD0113803);江苏省现代农机装备与技术示范推广项目(NJ2021-39)。

摘　　要：[目的]针对人工巡检哺乳期死亡仔猪费时费力且易引起母猪应激反应等问题,提出一种基于图像配准融合和改进YOLOv7的哺乳期死亡仔猪自动化检测方法。[方法]使用KAZE特征点匹配算法将可见光图像与热红外图像进行匹配,利用几何变换将配准图像空间对齐,通过Curvelet变换将待配准图像进行分解重构逆变为融合图像。以YOLOv7模型为基础,将SE注意力模块引入原始网络的Backbone部分,形成改进模型YOLOv7-SE,降低图像中低暗背景信息对目标识别的干扰,从而提升模型的检测性能。[结果]试验结果表明:模型在融合图像上的准确率、召回率与平均精度均值均高于可见光图像与热红外图像;与原始YOLOv7相比,YOLOv7-SE在准确率和召回率上分别提升3.2%和4.3%,平均单幅图片检测时间仅为6.8 ms。[结论]该模型可以实现养殖场场景下哺乳期死亡仔猪准确快速检测。[Objectives]A method for automated detection of lactating dead pigs based on image registration fusion and improved YOLOv7 was proposed to address the issues of time-consuming and labor-intensive manual inspection of lactating dead pigs,which can easily cause stress reactions in sows.[Methods]The KAZE feature point matching algorithm was used to match visible light images with thermal infrared images.The registered images spatially using geometric transformations was aligned,and the registered images were converted,fused,and inverted into fused images using the Curvelet transform in the frequency domain.Based on the YOLOv7 model,the SE attention module was introduced and embedded in the Backbone part of the original network to form an improved model YOLOv7-SE,which reduced the interference of low dark background information in the image on target recognition and improved the detection performance of the model.[Results]The experimental results showed that the accuracy,recall,and average accuracy of the model on fused images were higher than those on visible light and thermal infrared images.Compared with the original YOLOv7,YOLOv7-SE had improved accuracy and recall by 3.2%and 4.3%respectively,with an average single image detection time of only 6.8 ms.[Conclusions]This model can achieve accurate and rapid detection of dead piglets during lactation in breeding farm scenarios.

关 键 词：死亡仔猪 哺乳期 图像配准融合 YOLOv7-SE 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S828[自动化与计算机技术—计算机科学与技术]