Multi-Domain Image Super-Resolution Generative Adversarial Network for Low-Resolution Person Re-Identification

Person re-identification (ReID) is an important task in video surveillance application. To address the issue that various low-resolutions and scale mismatching always exist in the real world, a multi-domain image-to-image translation network, termed Multi-Domain image Super-Resolution Generative Adversarial Network (MSRGAN), is proposed to learn the mapping relationship between the various low-resolution domains and the high-resolution domain. MSRGAN can ensure that the transferred image has a similar resolution as in the target domain. It is also able to keep the identity information of images from low-resolution domain during the translation. In addition, a novel ReID model, termed CSA-ReID in which channel attention and spatial attention module are introduced, is designed to learn resolution-invariant deep representations. The proposed method achieves 90.7% rank-1 accuracy and 96.4% rank-5 accuracy on multiple low-resolutions Market-1501 dataset. The experimental results prove that the proposed method achieves promising generalization ability and accuracy compared with the state-of-the-art methods.