基于YOLO特征检测模型的非法溜宠物识别

Recognition of Illegal Pet Walking Based on YOLO Feature Detection Model

摘要: 非法溜宠物现象的识别属于细粒度图像分类的一种，针对复杂的生活环境，使用传统图像分类方法或者单纯的卷积神经网络来进行非法溜宠物现象的识别，会出现准确率偏低的情况。本文基于深度学习中实例分割方法来实现对溜宠物行为的识别，通过检测到的图像中目标的信息、目标与目标关系信息来实现对是否是非法溜宠物的现象判断。该方法是对2020年Glenn Jocher发表的实例分割模型YOLO (you only look once)的基础上进行改进，主要针对主干特征提取网络改为对细粒度图像更为友好的SENet特征网络。并将最深的stage部分的特征网络改为SPPCSPC模块用于优化特征提取精度。针对多种犬类和大量不同的现实环境的搭配来对整个网络进行训练。通过实际的公园活动场景中的识别表明，改进后的网络精度上变化不大，且很好地满足了实际生活中对于这种溜宠物现象的准确识别需要。

Abstract: The recognition of illegal pet walking is a kind of fine-grained image classification. For complex living environments, using traditional image classification methods or simple convolutional neural networks to recognize illegal pet walking will lead to low accuracy. This paper realizes the recognition of pet walking behavior based on the case segmentation method in depth learning, and judges whether it is illegal pet walking phenomenon by detecting the information of the target in the image and the relationship between the target and the target. This method is an improvement on the case segmentation model YOLO (you only look once) published by Glenn Jocher in 2020. It is mainly aimed at changing the backbone feature extraction network to the SENet feature network that is friendlier to fine grained images. The feature network of the deepest stage is changed into SPPCSPC module to optimize the feature extraction accuracy and train the whole network according to the combination of a variety of dogs and a large number of different real environments. The identification in the actual park activity scene shows that the accuracy of the improved network has little change, and meets the needs of accurate identification of this pet walking phenomenon in real life.