LocMix: local saliency-based data augmentation for image classification

Data augmentation is a crucial strategy to tackle issues like inadequate model robustness and a significant generalization gap. It is proven to combat overfitting, elevate deep neural network performance, and enhance generalization, particularly when data are limited. In recent years, mixed sample data augmentation (MSDA), including variants like Mixup and CutMix, has gained significant attention. However, these methods sometimes confound the network with misleading signals, limiting their effectiveness. In this context, we propose LocMix, an MSDA that aims to generate new training samples by prioritizing local saliency feature information and employing statistical data mixing. We achieve this by concealing salient regions with random masks and efficiently combining images through the optimization of local saliency information using transportation methods. Prioritizing the local features within an image allows LocMix to capture image details with greater accuracy and comprehensiveness, thereby enhancing the model's capacity to understand the target image. We conduct extensive validation of this approach on various challenging datasets. When applied to the training of the PreAct-ResNet18 model, our method yields notable improvements in accuracy. Specifically, on the CIFAR-10 dataset, we observe an impressive 1.71% accuracy enhancement. Similarly, on CIFAR-100, Tiny-ImageNet, ImageNet, and SVHN, we attain substantial accuracy improvements of 80.12%, 64.60%, 77.62%, and 97.12%, corresponding to improvements of 4.88%, 8.75%, 1.93%, and 0.57%, respectively. These experimental results plainly illustrate the effectiveness of our proposed method.