基于ECA-YOLO v5s网络的重度遮挡肉牛目标识别方法

肉牛目标检测和数量统计是精细化、自动化、智能化肉牛养殖要解决的关键问题，受肉牛个体颜色及纹理相近和遮挡等因素的影响，现有肉牛目标检测方法实用性较差。本研究基于YOLO v5s网络与通道信息注意力模块(ECABasicBlock),提出了一种融合通道信息的改进YOLO v5s网络(ECA-YOLO v5s),在YOLO v5s模型的骨干特征提取网络部分添加了3层通道信息注意力模块。ECA-YOLO v5s网络实现了重度遮挡环境下多目标肉牛的准确识别。对养殖场监控视频分帧得到的肉牛图像采用了一种基于结构相似性的冗余图像剔除方法以保证数据集质量。数据集制作完成后经过300次迭代训练，得到模型的精确率为89.8%,召回率为76.9%,全类平均精度均值为85.3%,检测速度为76.9 f/s,模型内存占用量为24 MB。与YOLO v5s模型相比，ECA-YOLO v5s的精确率、召回率和平均精度均值分别比YOLO v5s高1.0、0.8、2.2个百分点。为了验证不同注意力机制应用于YOLO v5s的性能差异，本研究对比了CBAM(Convolutional block attention mo...

Recognition Method of Heavily Occluded Beef Cattle Targets Based on ECA-YOLO v5s

Beef cattle target detection and quantity statistics are the first key problems should be solved in fine, automatic and intelligent beef cattle breeding. However, in the detection process, the existing beef cattle target detection methods cannot be applied to the actual beef cattle breeding because the beef cattle target colors are similar and there are severe occlusion with each other. Based on YOLO v5s and ECABasicBlock, a multi-target beef cattle detection method named ECA-YOLO v5s was proposed. The improvement method was to add three layers ECABasicBlock to the backbone feature extraction network of YOLO v5s model. YOLO v5s network, which integrated channel information, realized the accurate recognition of multi-target beef cattle in severe occlusion environment. A method of eliminating redundant images based on structural similarity was adopted to ensure the quality of beef cattle images. After labeling the beef cattle image obtained by framing the monitoring video of the farm, it was sent to ECA-YOLO v5s network for beef cattle target detection. After 300 iterations of training, the accuracy of the model was 89,8%, the recall rate was 76.9%, the mAP was 85.3%, the detection speed was 76.9f/s, and the model size was 24MB. Compared with YOLO v5s models, the precision value, recall value and mAP value of the ECA-YOLO v5s were 1.0 percentage points, 0.8 percentage points and 2.2 percentage points higher than those of YOLO v5s, respectively. Simultaneously, the performance differences of different attention mechanisms applied to YOLO v5s were compared, and the four attention mechanisms of CBAM, CA, SE and ECA were compared. By comparison, the mAP value of ECA-YOLO v5s was 0.5 percentage points, 0.6 percentage points and 0.2 percentage points higher than that of CBAM, CA and SE, respectively. It could be seen that the network effect of integrating ECA module was the best. The detection results of different occlusion and illumination conditions were analyzed and discussed. The results showed that ECA-YOLO v5s network can accurately and quickly detect beef targets with different occlusion and illumination conditions. The model had high robustness and small model, which was convenient for the migration and application of the model, and the research result can provide necessary technical support for the research of beef cattle target detection and pledge supervision.