Blend AutoAugment: Automatic Data Augmentation for Image Classification Using Linear Blending

Abstract

Data augmentation improves machine learning model performance by diversifying training data. Initially, manual selection of augmentation techniques was required; however, AutoAugment automated this process. However, recent studies identified limits in the search space of AutoAugment, which hindered performance gains. In this study, we introduced a novel method called Blend AutoAugment that utilizes linear blending with random weights to enhance the performance of image classification in deep learning. Our method consists of two stages: 1) policy search and 2) image blending. In the policy search stage, optimal policies are determined by employing reinforcement learning to explore a search space, comprising the types of operations, their application probabilities, and their magnitudes. In the image blending stage, multiple optimal policies discovered in the previous stage are selected to generate augmented images, and the generated images are blended using random weights. Performance comparison experiments were carried out using the Wide-ResNet40-2 and Wide-ResNet28-10 models on the CIFAR-10, CIFAR-100, CIFAR-10-C, and CIFAR-100-C datasets. In addition, the proposed method was compared with existing methods. Notably, the proposed method exhibited higher accuracy and robustness than those of existing methods. Furthermore, it was applicable to various existing automatic data augmentation methods. In summary, this study introduced an advanced automatic data augmentation method that expands the previously limited search space and utilizes image blending to achieve better accuracy and robustness than conventional methods. Our source code is available at https://github.com/Sarrzae/BAA.

Keywords

• Data augmentation
• deep learning
• image classification
• image processing