Superpixel Mixing: A Data Augmentation Technique For Robust Deep Visual Recognition Models

Data augmentation can mitigate overfitting problems in data exploration without increasing the size of the model. Existing cutmix-based data augmentation has been proven to significantly enhance deep learning performance. However, many existing methods overlook the discriminative local context of the image and rely on ad hoc regions consisting of square or rectangular local regions, resulting in the loss of complete semantic object parts. In this work, we propose a superpixelwise local-context-aware efficient image mixing approach for data augmentation, aiming to overcome the limitations previously mentioned. Our approach only requires one forward propagation using a superpixel attention-based label mixing with lower computational complexity. The model is trained using a combination of a global classification of the mixed (augmented) image loss, a superpixel-wise weighted local classification loss, and a superpixel-based weighted contrastive learning loss. The last two losses are based on the superpixel-aware attentive embeddings. Thus, the resulting deep encoder can learn both local and global features of the images, capturing object-part local context information. Experiments on diverse benchmarks, such as ImageNet-1K and CUB-200-2011, indicate that the proposed method out-performs many augmentation methods for visual recognition. We have not only demonstrated its effectiveness on CNN models, but also on transformer models.