Self-transformation of encoded subregion image data augmentation for better classification

Abstract

Data augmentation has been proven to be an effective regularization strategy that can reduce over-fitting risks in deep learning models. Erasure based data augmentation is one of the most advanced solutions; however, random region erasure inevitably leads to excessive loss of object information and the introduction of a large amount of negative noise. In this work, a data augmentation method based on self-transformation of encoded image subregions is proposed. This method first designs an encoding mechanism based on random strategy to determine the object sub-regions of the transformation. Then, self-morphing strategies such as random cropping, rotation, and channel transformation are applied to the sub-regions to maintain the original feature information conditions and increase the local variance characteristics of the samples. Finally, we perform erasure on the non-discriminative triangular region generated after rotation of the object region. The proposed method is easy to implement, and a large number of experiments were conducted on challenging datasets. The algorithm achieved significant improvements in classification performance. On the CIFAR-100 dataset, the top-1 error rate decreased by 2.83%. For object detection, the average detection accuracy on the PASCAL VOC dataset increased from 79.83% to 82.15%, with the highest average accuracy improvement of 5.32% for a single category.