Fast Recognition of Pig Faces Based on Improved Yolov3

Abstract

Abstract The traditional ear labels are easy to get off and cause infection in the intelligent managements of live pigs. Therefore, a noninvasive pig face recognition method based on improved YOLOv3 was designed to carry out simultaneous recognition of multiple live pigs.The YOLOv3 model was used to introduce the dense network into the Darknet53 feature extractor, forming a new backbone network in combination with down-sampling, and the improved SPP unit was added to the YOLOv3 model, constructing the YOLOv3_DB_SPP model. The pig face data set used in the test was divided into 10 categories. After data enhancement, the number of samples was 8 512, and the ratio of training set to test set was about 9:1. The test results showed that: 1) Under different classification probability thresholds, the mean average precision of the YOLOv3_DB_SPP model for detecting pig face data set was higher than that of the YOLOv3 model; 2) When the IOU threshold was 0.5 and the classification probability threshold was 0.1, the mean average precision of YOLOv3_DB_SPP model was 9.87% higher than that of the YOLOv3 model; 3) When detecting long-distance covered small target samples, the mean average precision was also higher in the YOLOv3_DB_SPP model. Thus, the YOLOv3_DB_SPP model improved the feature extraction ability of the basic feature extractor and the precision of the detector.