基于图像语义分割的车辆重识别

Vehicle re-identification based on image semantic segmentation

针对同一款式不同车辆在外观上具有较大相似性,且从不同视点车型外观的变化较大等问题,结合全局特征和局部特征,提出一个框架,即图像语义分割重识别方法。对车辆进行语义分割,去除车辆的背景信息,提取出车辆局部分块,得到3个局部特征;结合整车的全局特征,联合分类损失和三重态损失对特征模块进行训练,融合局部损失和全局损失,对网络进行优化。利用该方法解决三重态训练中全局信息不能提供局部信息的问题,减少背景噪声对局部特征的影响。实验结果表明了该方法的有效性。

Different vehicles of the same style have relatively large similarities in appearance,and the appcarance of different models from different viewpoints varies greatly.Combining global features and local features,a framework was proposed,that was,the image semantic segmentation and re-recognition method.The vehicle was semantically segmented,the background information of the vehicle was removed,and the local parts of the vehicle were extracted to obtain three local features.The global characteristics of the whole vehicle were combined,classification loss and triplet loss were jointed to train the feature module,and the local loss and global loss were merged to optimize the network.Using this method not only solves the problem that global information cannot provide local information in triplet training,but reduces the influence of background noise on local features.Experimental results show the effectiveness of the proposed method.