基于多尺度生成对抗网络的遮挡行人重识别方法

行人重识别是指在多个非重叠摄像头拍摄的场景下，给定一幅查询行人图像，从大规模行人图像库中检索出具有相同身份的行人图像，是一类特殊的图像检索任务.随着深度学习的不断发展，行人重识别方法的性能得到了显著提升.但是行人重识别在实际应用中经常遭遇遮挡问题（例如背景遮挡、行人互相遮挡等）.由于遮挡图像不仅丢失了部分目标信息，而且引入了额外的干扰，使得现有方法往往难以学习到鲁棒的特征表示，从而导致识别性能严重下降.最近，生成对抗网络在各类计算机视觉任务上展现出强大的图像生成能力.受到生成对抗网络的启发，提出了一种基于多尺度生成对抗网络的遮挡行人重识别方法.首先，利用成对的遮挡图像和非遮挡图像训练一个多尺度生成器和一个判别器.多尺度生成器能够对随机遮挡区域进行去遮挡操作，生成高质量的重构图；而判别器能够区分输入图像是真实图像还是生成图像.其次，利用训练好的多尺度生成器，生成去除随机遮挡的训练图像，添加到原始训练图像集，用于增加训练样本的多样性.最后，基于此扩充训练图像集，训练分类识别模型，有效地提高模型在测试图像集上的泛化性.在多个有挑战性的行人重识别数据集上的实验结果，验证了所提出方法的有效性.

Multi-scale Generative Adversarial Network for Person Re-identification under Occlusion

Person re-identification (ReID) refers to the task of retrieving a given probe pedestrian image from a large-scale gallery collected by multiple non-overlapping cameras, which belongs to a specific task of image retrieval. With the development of deep learning, the performance of person ReID has been significantly improved. However, in practical applications, person ReID usually suffers from the problem of occlusion (such as background occlusion, pedestrian occlusion). The occluded image not only loses partial target information, but also introduces additional interference, which makes the deep neural network difficult to learn robust feature representations and seriously degrades the performance of person ReID. Recently, generative adversarial network (GAN) has shown the powerful image generation ability on various computer vision tasks. Inspried by GAN, we propose a person ReID method under occlusion based on multi-scale GAN. Firstly, we use the paired occluded images and unoccluded images to train a multi-scale generator and a discriminator. The multi-scale generator can restore the lost information for randomly occluded areas and generate high-quality reconstructed images; while the discriminator can distinguish whether the input image is a real image or a generated image. Then, we use the trained multi-scale generator to generate the de-occluded images. Adding these de-occluded images to the original training image set can increase the diversity of training samples. Finally, we train a classification network based on the augmented training image set, which effectively improves the generalization capability of the trained model on the testing image set. Experimental results on several challenging person ReID datasets demonstrate the effectiveness of our proposed method.