Occluded pedestrian re-identification primarily focuses on how to effectively extract feature representations from the non-occluded regions when parts of a pedestrian's body are obstructed in surveillance footage, and how to perform similarity analysis of the occluded features in a reasonable manner. In this paper, we used Generative Adversarial Network (GAN) to remove occlusion from interest objects with occlusion information in video sequences, and designing a multi-scale generator to further enhance the data generation capability of the network, thereby reducing the interference of occlusion during the feature representation learning process of interest objects. Experimental results on the occluded pedestrian re-identification dataset show that, compared to mainstream occlusion recognition models, our method achieves better recognition performance and reaches state-of-the-art recognition accuracy.