Aiming at the problem of person image style difference caused by cross-camera shooting in the process of person re-identification, this paper proposes a learning framework based on a cyclic vector quantization generative adversarial network (CVQGAN) and a self-calibrated convolution module. First of all, this paper designs a discrete vector quantization module, which is introduced into the process from encoding to decoding of the generator. The discrete vector in the vector quantization space is used to solve the problem that the original generator produces noisy pseudo images, therefore generating higher quality style conversion images. Then, the self-calibration convolution module is integrated into the convolution layer of the Resnet50 backbone network, and the multi-branch network structure is used to perform different convolution operations on each branch, so as to obtain features with stronger characterization ability and further solve the problem of style differences of the same pedestrian under different cameras. The proposed algorithm is validated by experiments on Market1501 and DukeMTMC-reID datasets, and the results show that the proposed algorithm can effectively improve the accuracy and robustness of person re-identification.