Abstract: In the ex situ conservation and population management of giant pandas (Ailuropoda melanoleuca), timely and rapid individual identification and behavior monitoring play a crucial role in health supervision. The health status of captive giant pandas is usually observed by specialized breeders, which have high labor costs, low efficiency, and lack of timeliness. Therefore, efficient and cheap image-based identity recognition and behavior analysis has become a new development trend. Existing studies have proposed to realize it by detecting and analyzing facial images of giant pandas. However, this type of method still has the problem of insufficient detection accuracy, which makes it difficult to improve the recognition performance. We propose a dual model fusion method based on YOLOv3 and Mask R-CNN to achieve the segmentation and accurate detection of giant panda head images. It includes three parts:YOLOv3 completes head detection, Mask R-CNN completes panda contour segmentation, and then fuses the outputs of the two models. Experimental results show that the accuracy of head detection, giant panda and head contour segmentation are 82. 6%, 95. 2%, and 87. 1%, respectively. Our method has high detection and segmentation accuracy for the giant panda head images. It provides help for individual identification and gender classification of giant pandas and provides technical reference for behavior analysis.

Key words: Giant panda, YOLOv3, Mask R-CNN, Dual model fusion, Image segmentation

摘要： 在大熊猫(Ailuropoda melanoleuca)的迁地保护和种群饲养管理中，及时、快速地进行个体识别和行为监测，对其健康管理具有至关重要的作用。圈养大熊猫健康状况通常由专门的饲养人员肉眼观测，人力成本高、效率低并且缺乏时效性。基于图像的动物个体识别与行为分析技术效率高、时间成本低，已经成为新的监测发展趋势。已有研究提出，通过大熊猫面部图像的检测和分析，可实现个体识别和行为分类。但该方法依然存在检测精度不足导致识别准确率难以提升的问题。本文提出一种基于YOLOv3和Mask R-CNN的双模型融合方法，实现了大熊猫头部图像分割和精准检测。包含3个部分：YOLOv3完成头部检测，Mask R-CNN完成大熊猫轮廓分割，然后将两个模型的输出进行交并比融合。结果显示，头部检测准确率为82.6%，大熊猫轮廓分割准确率为95.2%，总体头部轮廓分割准确率为87.1%。该方法对大熊猫头部图像的识别率和分割准确率高，为大熊猫的个体识别、性别分类提供了帮助，为行为分析提供了技术参考。

关键词: 大熊猫, YOLOv3, Mask R-CNN, 双模型融合, 图像分割