DWT+DWT: Deep Learning Domain Generalization Techniques Using Discrete Wavelet Transform with Deep Whitening Transform

Recently, there is a growing demand for a deep learning framework with robust generalization performance in real-world domains, such as an autonomous driving environment. The existing domain generalization methodologies for convolutional neural networks have been designed to actively utilize the feature map with the generative model or normalization techniques to distinguish domain-specific information. However, augmented images are essential for measuring style sensitivity. This study shows that style information can be extracted from an original image through color space separation and frequency decomposition without a separate augmented image. Therefore, it can be used as a method independent of existing network models. The proposed method shows an mIoU improvement by 1.54% compared to the existing method in the semantic segmentation model trained using urban scene datasets.

I. Introduction

Modern computer vision tasks are constantly evolving with the deep learning frameworks inspired by various human vision systems [1]–[4]. However, the performance of each task is highly dependent on the training data, and the use of limited data hinders the performance improvement of vision tasks [5]. For generalization of the models, early deep learning techniques suggested a solution for the overfitting problem on datasets with almost similar data distributions (i.e., same domain) [6]. However, with the development of deep learning models, the demand for out-of-distribution (OOD), called domain overfitting, has emerged (e.g., domain shift) [7].