A Simplified Feature Alignment Strategy for Image Classification Across Domains

Abstract

Recently, in deep learning research, the importance of domain generalization (DG) for unseen domains has been emphasized. Most of the baseline methodologies for this focus on generating adversarial representations or separating content and style information from intermediate features for learning. However, these approaches inevitably increase the time complexity for both training and inference. In this study, we propose an approach to improve DG performance without excessive bottleneck points. We suggest an auxiliary network structure that places a mapping layer for feature alignment after the stem layer, a generative model based on an adaptive instance normalization that can adjust mean and standard deviation. This structure consistently adjusts the output feature maps of the stem layer to follow a Gaussian distribution regardless of the domain used as the input image. Moreover, both training and inference are possible without iterative routines, making their complexity nearly identical to training without the DG strategies. Experimental results show that our model outperforms the existing DG baseline with the highest performance in image classification tasks by an average accuracy of 0.71% higher on the PACS benchmarking dataset.