PatchMix: patch-level mixup for data augmentation in convolutional neural networks

Convolutional neural networks (CNNs) have demonstrated impressive performance in fitting data distribution. However, due to the complexity in learning intricate features from data, networks usually experience overfitting during the training. To address this issue, many data augmentation techniques have been proposed to expand the representation of the training data, thereby improving the generalization ability of CNNs. Inspired by jigsaw puzzles, we propose PatchMix, a novel mixup-based augmentation method that applies mixup to patches within an image to extract abundant and varied information from it. At the input level of CNNs, PatchMix can generate a multitude of reliable training samples through an integrated and controllable approach that encompasses cropping, combining, blurring, and more. Additionally, we propose PatchMix-R to enhance the robustness of the model against perturbations by processing adjacent pixels. Easy to implement, our methods can be integrated with most CNN-based classification models and combined with varying data augmentation techniques. The experiments show that PatchMix and PatchMix-R consistently outperform other state-of-the-art methods in terms of accuracy and robustness. Class activation mappings of the trained model are also investigated to visualize the effectiveness of our approach.