基于动态前景聚焦与伪孪生网络的跨分辨率行人重识别

Cross resolution person re-identification based on dynamic foreground focus and pseudo siamese network

针对跨分辨率场景下行人图像存在场景复杂、重建图像特征提取效果差等问题,提出基于动态前景聚焦与伪孪生网络的跨分辨率行人重识别算法.该算法在超分辨率重建网络中嵌入动态前景聚焦模块,利用全卷积自动编码器提取目标行人特征,通过高斯掩码对网络进行空间引导,从而使判别特征聚焦在前景上;并经过动态感知模块自动捕获前景的重要特征.又通过构建多粒度相互协同的伪孪生网络,实现判别特征的精细化识别.最后,所提算法在跨分辨率数据集MLR-Market-1501,MLR-DukeMTMC-ReID和CAVIAR上进行实验,Rank-1精度分别达到了91.3%,83.4%和48.5%,证明了所提算法对跨分辨率行人重识别任务的有效性.

Aiming at the problems of pedestrian images in cross resolution scenes,such as unbalanced foreground background and poor feature extraction of reconstructed images,a cross resolution person re-identification recognition algorithm based on dynamic foreground focus and pseudo siamese network was proposed.The algorithm embeds a dynamic foreground focus module in the super-resolution reconstruction network,amplifies the pedestrian foreground with a full convolution automatic encoder,and guides the network through a Gaussian mask to focus the discriminant features on the foreground.The dynamic perception module automatically captured the important features of the foreground by constructing the pseudo siamese network of coarse and fine granularity,and the fine recognition of discriminant features was realized.Finally,the proposed algorithm was tested on the cross resolution dataset MLR-Market-1501,MLR DukeMTMC-ReID and CAVIAR,and the accuracy of Rank-1 was 91.3%,83.4%and 48.5%,respectively,which proves the effectiveness of the proposed algorithm for cross resolution person re-identification task.