基于改进YOLO v4的肉鸽行为检测模型研究

Pigeon Behavior Detection Model Based on Improved YOLO v4

肉鸽行为表现与鸽舍环境舒适度和肉鸽健康状况密切相关。为实现肉鸽行为精准检测、及时掌握肉鸽健康状况,提出了基于改进YOLO v4模型的肉鸽行为检测方法。由于肉鸽社交等行为特征相似性程度高,为了在复杂环境下准确识别肉鸽行为,本文采用自适应空间特征融合(Adaptively spatial feature fusion,ASFF)模块改进YOLO v4模型,在特征金字塔网络中增加ASFF模块,根据特征权值自适应融合多层特征,充分利用不同尺度特征信息,并且ASFF模块能有效过滤空间冲突信息、抑制反向梯度不一致问题、改善特征比例不变性以及降低推理开销。基于多时段的肉鸽清洁和社交行为数据集,自制5类肉鸽行为图像数据库,采用OpenCV工具进行模糊、亮度、水雾和噪声等处理扩充图像数据集(共10320幅图像),增加数据多样性和模拟不同识别场景,提升模型泛化能力。本文按照比例8∶2划分训练集和验证集,训练总共迭代300个周期,对不同时段、角度、尺寸的肉鸽数据集进行检测。检测结果表明,在阈值0.50和0.75时YOLO v4 ASFF检测精度比YOLO v4的mAP50和mAP75提高14.73、14.97个百分点。对比Faster R CNN、SSD、YOLO v3、YOLO v5和CenterNet模型验证本文模型检测性能,在测试集中mAP50分别提高13.98、14.00、18.63、14.16、10.87个百分点。视频检测速度为8.1 f/s,在推理速度相当情况下,本文改进模型识别准确率更高,复杂环境泛化能力更强,且对相似度高的行为误检和漏检情况更少,可为智能化肉鸽养殖和科学管理提供技术参考。

Pigeon whole behavior is closely related to the loft environmental comfort and pigeon whole health.For human observation and recording the pigeon whole behavior is time⁃consuming,sampling limited,subjective and other issues,to timely meet the pigeon whole precision detection and pigeon whole behavior and health,based on the YOLO v4 pigeon whole behavior detection method was proposed.In this method,CSPDarkNet53 was used as the Backbone network to extract feature maps covering shallow semantic information of pigeons,and then PANet was used to transfer the bottom features and stack features to the top.Aiming at the high similarity degree of pigeon social behavior features,in order to achieve accurate identification of pigeon behavior in complex environment.The adaptively spatial feature fusion(ASFF)module was adopted to improve the YOLO v4 model,and the ASFF module was added to the feature pyramid network,which can adaptively fuse multi⁃layer features according to the feature weights and make full use of the features information of different scales.Moreover,ASFF can effectively filter spatial conflict information and suppress reverse gradient inconsistency,improve feature proportion invariance and reduce inference overhead.Based on the cleaning and social behaviors of meat pigeons in multiple periods,a database of five kinds of meat pigeon behavior images was made.OpenCV tool was used to process blur,brightness,haze and noise to expand the image data set(totally 10320 images),increase data diversity and simulate different recognition scenes,and improve the generalization ability of the model.A 8∶2 ratio was used to divide the training and validation sets.The training iterated 300 epochs in total,and the detection was carried out through meat pigeon data sets of different time periods,angles and sizes.The detection results showed that the detection accuracy of improved YOLO v4 ASFF model was 14.73 percentage points and 14.97 percentage points higher than that of mAP50 and mAP75 of original YOLO v4 model at the threshold of 0.50 and 0.75.Compared with Faster R CNN,SSD,YOLO v3,YOLO v5 and CenterNet model,mAP50 of the YOLO v4 ASFF was improved by 13.98 percentage points,14.00 percentage points,18.63 percentage points,14.16 percentage points and 10.87 percentage points in test set,respectively.The video detection speed was 8.1 f/s,and the improved model had higher recognition accuracy under the condition of the same inference speed,strong generalization ability in complex environment,and less misdetection and omission of behaviors with high similarity.The research on meat pigeon behavior detection can provide technical reference for intelligent meat pigeon breeding and scientific management.