基于多辅助分支深度网络的行人再识别

Person Re-Identification based on Deep Learning with Multi-Auxiliary Branches

利用广泛遍布的摄像机进行再识别,一直是近年的研究热点和难点,在智能监控、智能安防等领域应用广泛。随着深度学习在图像识别领域的重大突破,不断提出的网络结构诸如AlexNet、VGGNet、ResNet等,通过数十层甚至上百层的卷积神经网络,提取出更加鲁棒的特征进行分类,其性能已经远远超出传统手工设计特征的再识别方法。针对深度网络结构只利用多层卷积后输出的高层语义特征而忽视中间层特征的问题,提出基于多辅助分支神经网络的行人再识别算法。该模型能够充分利用中间层特征所保留的信息,弥补由于多层卷积造成的图片部分信息丢失。最后,通过公开数据集测试网络性能,证明了该算法能明显提升行人再识别的准确率。

Person re-identification （re-id） based on extensively-distributed cameras is widely applied in intelligent monitoring, intelligent security and other fields, and always regarded as the research hotspot and difficulty. In recent years, with the significant breakthrough of deep learning in the field of image recognition, the ever-proposing network structures such as AlexNet, VGGNet, ResNet, etc., can extract more robust features for classification through dozens or even hundreds of convolutional neural networks, and their performance is far-exceeded the re-identification method of traditional manual design features. Aiming at the problem that the deep network structure only uses the high-level semantic features of multi-level convolution output and neglects the characteristics of the middle layer, a pedestrian re-recognition algorithm based on multi auxiliary branches neural network is proposed. The model can make full use of the information retained by the middle layer features to compensate the loss of image information resulted from multi-layer convolution. Finally, the experimental test with public dataset on the network performance indicates that this algorithm could significantly improve the accuracy of person re-id.