基于EfficientDet的围产期母猪姿态识别

Recognition of Perinatal Sows’ Posture Based on EfficientDet Algorithm

围产期母猪母性行为直接影响仔猪的成活率,母猪姿态是其母性行为和筑巢行为的重要表现。针对目前对围产期母猪姿态转换主要依靠人工巡检,费时耗力且主观性强等问题,采集了24头母猪的视频数据并对数据进行预处理,利用EfficientDet网络对产床内母猪图像进行深层次特征提取,实现了母猪站、坐、胸卧、侧卧姿态及其侧卧方向(乳房面向仔猪保温箱、乳房背对仔猪保温箱)的准确识别。结果表明:该模型识别平均精度均值(mAP)达93.97%,对图像的检测速度达26.2 f/s,对视频的检测速度达10.66 f/s。通过对母猪产前及产后24 h的姿态进行分析,母猪产前表现出显著的筑巢行为,姿态转换频率显著提高(P<0.001);母猪产后侧卧时间显著增加,母猪侧卧时长与仔猪窝均质量呈正相关关系;根据母猪侧卧方向的偏好性进行分组比较,母猪偏向于将乳房面向保温箱侧卧的小组,仔猪断奶成活率更高。

The maternal behavior of sows in the perinatal period directly affects the survival rate of piglets, and the posture of the sows is an important manifestation of its maternal behavior and nesting behavior. Current perinatal sow’ posture transformation mainly relies on manual observation, which is time-consuming, labor-intensive and subjective. Aiming to recognize sows' stand, sit, sternum, lateral and lateral directions totally 24 sows’ video were collected and the data pre-processing research was performed, and the EfficientDet network was used to perform deep feature extraction on the image of sows in the farrowing bed. The breast facing the heat lamp for piglet and the breast averting the heat lamp were included. The results showed that the average recognition accuracy (mAP) of the algorithm was 93.97%, the detection speed of pictures was 26.2 f/s, and the detection speed of videos was 10.66 f/s. By analyzing the postpartum of sows before and 24 hours after delivery, sows showed significant nesting behaviors before delivery, and the frequency of posture changes was increased obviously(P<0.001). Sows' lateral time after delivery was increased significantly, there was a positive correlation between the length of lateral time and the average weight of the piglets. The sows were grouped according to the preference of the lateral direction. When the sows tended to lie with breast facing the heat lamp, the piglets’ survival rate was higher.