This paper proposes a semi-supervised learning method that uses generative adversarial networks (GANs) to generate unlabeled samples, which are then integrated into a convolutional neural network (CNN) for person re-identification (re-ID). The main contribution is the label smoothing regularization for outliers (LSRO), which assigns a uniform label distribution to the unlabeled images, regularizing the supervised model and improving the baseline. The method is evaluated on three large-scale datasets: Market-1501, CUHK03, and DukeMTMC-reID, achieving improvements of +4.37%, +1.6%, and +2.46% in rank-1 precision over the baseline CNN, respectively. Additionally, the method is applied to fine-grained bird recognition, achieving a +0.6% improvement over a strong baseline. The GAN-generated data enhances the discriminative ability of the learned CNN embeddings, and the LSRO method is shown to be superior to alternative approaches such as "All in one" and "Pseudo label". The results demonstrate that the proposed method effectively improves performance in both re-ID and fine-grained recognition tasks. The method is implemented using a DCGAN model for sample generation and a ResNet model for representation learning. The code is available at https://github.com/layumi/Person-reID_GAN.This paper proposes a semi-supervised learning method that uses generative adversarial networks (GANs) to generate unlabeled samples, which are then integrated into a convolutional neural network (CNN) for person re-identification (re-ID). The main contribution is the label smoothing regularization for outliers (LSRO), which assigns a uniform label distribution to the unlabeled images, regularizing the supervised model and improving the baseline. The method is evaluated on three large-scale datasets: Market-1501, CUHK03, and DukeMTMC-reID, achieving improvements of +4.37%, +1.6%, and +2.46% in rank-1 precision over the baseline CNN, respectively. Additionally, the method is applied to fine-grained bird recognition, achieving a +0.6% improvement over a strong baseline. The GAN-generated data enhances the discriminative ability of the learned CNN embeddings, and the LSRO method is shown to be superior to alternative approaches such as "All in one" and "Pseudo label". The results demonstrate that the proposed method effectively improves performance in both re-ID and fine-grained recognition tasks. The method is implemented using a DCGAN model for sample generation and a ResNet model for representation learning. The code is available at https://github.com/layumi/Person-reID_GAN.