A Cattle Detection Algorithm Based on ResMO-Dense-YOLO
Cattle target detection is a prerequisite for individual registration and recognition of cattle based on deep learning.The differences in lighting,color and breed in different actual scenarios make low-level features of cattle images diverse,while semantic information in high-level features cannot fully match the diverse low-level features,resulting in poor detection accuracy.In order to solve the problem of insufficient high-level feature semantics of the detection model,this paper designs a new cattle feature extraction backbone network ResMO Backbone and feature fusion network Dense Neck,and proposes a cattle detection algorithm based on ResMO Sense YOLO.In the backbone network,the ResMO module(ResBlock MHSA ODConv)is used to focus on the characteristics of cattle high-level features at multi-semantic level to enrich semantic information,and the SPPF structure and multi-layer convolution structure are combined to expand the receptive field,so that the model can better extract cattle high-level features;then,a feature pyramid based on DenseBlock and a feature fusion network cascaded with a path aggregation network based on DenseBlock are proposed,which utilize the feature reuse feature of DenseBlock and combine the multi-scale fusion feature of the feature pyramid and path aggregation network to further integrate the low-level feature position information and high-level feature semantic information of cattle,improving model detection accuracy.Compared with the FLYOLOv3,SSD and YOLOv5s,the model in this paper shows an average accuracy improvement of 40.1%,30.3%,and 4.0%in the data sets of cow channels,cow sheds,and beef sheds collected in the laboratory.The recall rate increased by 34.9%,23.1%,and 6.8%,respectively,and the mAP increased by 49.2%,35.3%,and 5.0%.
NETL
NSTL
万方数据
