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

In our study, we have explored the influence of panoramic images and ordinary images on the performance of Siberian crane detection, and compared the detection accuracy under different networks based on YOLOv5, to get fine and high-quality datasets and select the proper model for Serbian crane detection. The results show that (i) Training datasets from the internet and ordinary field photos can achieve a better detection performance than other training datasets, and Training datasets from panoramic images only show low accuracy due to Siberian crane's alertness and mosaic data enhancement method adopted in YOLOv5, which reduced the size of a small target. (ii) when the iteration times reach 40000, the YOLOv5 model can completely converge, and the mAP value reached 81.4%, total loss value 0.0357; (iii) With increasing the width and depth of layer in YOLOv5, the value of mAP show a growth trend, however the FPS show an opposite trend; (iv) through verification, we found that the model can also have an effectively performance of detection in the complex environments, such as multi-objective small objects and occlusions, the color similarity between target and background, different dynamic activities including flying, falling, foraging, playing, etc.