Optimized Detection Method for Siberian Crane (Grus leucogeranus) Based on Yolov5

EasyChair Preprint 5303

Abstract

In our study, we have compared the detection accuracy of the YOLOv5 model based on different training datasets and iteration times, to select the optimal dataset for the following research and build a  YOLOv5-based optimization detection model of Siberian crane by using deep learning method and training datasets collected by network crawling, panoramic cameras and SLR cameras. The results show that， (i) the mAP values of the six training datasets are Train_4 > Train_2 > Train_1 > Train_3 in turn. The mAP value of Train_4 was biggest, reaching to 90.9%, which was much higher than other training dataset, and indicate that the model detection accuracy of the training dataset mixed with ordinary field photos and network photos was much higher than that of the training dataset from a single source image; (ii)  when the iteration times  reach to 40000times,  YOLOv5 model can completely converge, and mAP value reached to 81.3%, total loss value 0.0357；(iii) According to the result of the existing model test, we found that the model can also have an effectively performance of detection in the complex environments, such as pictures exist the problems of multi-objective small objects and occlusions, similarity in color between the object and background, and Siberian cranes are in different activities such as flying, falling, foraging, playing etc.

Keyphrases: Siberian crane, YOLOv5, deep learning, object detection

@booklet{EasyChair:5303,
  author    = {Linlong Wang and Huaiqing Zhang and Tingdong Yang and Jing Zhang and Zeyu Cui and Nianfu Zhu and Yang Liu and Yuanqing Zuo},
  title     = {Optimized Detection Method for Siberian Crane (Grus leucogeranus) Based on Yolov5},
  howpublished = {EasyChair Preprint 5303},
  year      = {EasyChair, 2021}}