关键词: 关键词: 长尾数据, 目标检测, 两阶段学习, 重加权, YOLOv4-Tiny

Abstract: Abstract: Wild animal target detection based on infrared camera image is conducive to the research and protection of wild animals. Because of the large difference in the number of different species of wildlife, there is a long tail data problem with uneven distribution of species number in the wildlife data set collected by infrared camera. This problem affects the overall performance improvement of the target detection neural network model. In order to solved the problem of low accuracy of target detection caused by long tail data of wild animals, a method based on two-stage learning and re-weighting for long tail data was proposed, and the method was applied to wild animal target detection based on YOLOv4-Tiny network. Firstly, a new wildlife data set was constructed, which had obvious long tail data characteristics. Secondly, a two-stage method based on transfer learning was used to train the neural network. In the first stage, the classification loss function was trained without weighting. In the second stage, two improved re-weighting methods were proposed, and the weights obtained in the first stage were used as the pre-training weight for re-weighting training. Finally, the wild animal test set was used to tested. Experimental results show that the proposed methods achieved 61.18% mAP, which is 5.1% higher than the no-weighted method. It shows that the proposed method can improve the target detection performance of YOLOv4-Tiny network for wildlife data sets with long tail data characteristics.

Key words: long-tail data, object detection, two-stage learning, re-weighting, YOLOv4-Tiny