This paper introduces Random Erasing, a new data augmentation method for training convolutional neural networks (CNNs). Random Erasing randomly selects a rectangular region in an image and erases its pixels with random values, generating images with varying levels of occlusion. This method reduces overfitting and improves the model's robustness to occlusion. It is parameter-free, easy to implement, and can be integrated with most CNN-based recognition models. Random Erasing is complementary to common data augmentation techniques like random cropping and flipping, and consistently improves performance on image classification, object detection, and person re-identification. The method is also robust to partially occluded samples and outperforms baseline models on the CIFAR-10 dataset. The paper evaluates Random Erasing on three image classification datasets: CIFAR-10, CIFAR-100, and Fashion-MNIST. Results show that models trained with Random Erasing achieve significant improvements in accuracy. For example, on CIFAR-10, the method improves accuracy by 0.49% using ResNet-110 and 0.33% using ResNet-110-PreAct. On Fashion-MNIST, it reduces the top-1 error rate from 4.01% to 3.65%. The method is also effective for object detection and person re-identification, achieving competitive results on benchmark datasets like PASCAL VOC 2007, Market-1501, DukeMTMC-reID, and CUHK03. It outperforms existing methods in these tasks and is easy to implement without requiring parameter learning. The method is shown to be robust to occlusion and improves the generalization ability of CNNs. The paper concludes that Random Erasing is a simple yet effective data augmentation technique that enhances the performance of CNNs on various recognition tasks.This paper introduces Random Erasing, a new data augmentation method for training convolutional neural networks (CNNs). Random Erasing randomly selects a rectangular region in an image and erases its pixels with random values, generating images with varying levels of occlusion. This method reduces overfitting and improves the model's robustness to occlusion. It is parameter-free, easy to implement, and can be integrated with most CNN-based recognition models. Random Erasing is complementary to common data augmentation techniques like random cropping and flipping, and consistently improves performance on image classification, object detection, and person re-identification. The method is also robust to partially occluded samples and outperforms baseline models on the CIFAR-10 dataset. The paper evaluates Random Erasing on three image classification datasets: CIFAR-10, CIFAR-100, and Fashion-MNIST. Results show that models trained with Random Erasing achieve significant improvements in accuracy. For example, on CIFAR-10, the method improves accuracy by 0.49% using ResNet-110 and 0.33% using ResNet-110-PreAct. On Fashion-MNIST, it reduces the top-1 error rate from 4.01% to 3.65%. The method is also effective for object detection and person re-identification, achieving competitive results on benchmark datasets like PASCAL VOC 2007, Market-1501, DukeMTMC-reID, and CUHK03. It outperforms existing methods in these tasks and is easy to implement without requiring parameter learning. The method is shown to be robust to occlusion and improves the generalization ability of CNNs. The paper concludes that Random Erasing is a simple yet effective data augmentation technique that enhances the performance of CNNs on various recognition tasks.