BDIS: Balanced Training Architecture for Dual Image Scaler Using Origin Referenceable Losses

Abstract

Deep neural network (DNN)-based research on image scaling has mostly focused on super-resolution (SR) rather than image downscaling. Specifically, most existing DNN-based methods for image downscaling are used as auxiliary modules to improve the quality of super-resolved images. In rare cases, DNN-based methods consider the image downscaling as an important task as SR to increase the quality of downscaled images. In these methods, when setting the loss function for the training of the downscaling, the downscaled images generated by bicubic or bilinear interpolation are used as the ground-truth. As a result, the downscaled image by these methods cannot significantly differ from that by simple interpolation. In addition, these DNN-based methods with SR and downscaling modules have an imbalanced training architecture, which leads to biased training. To resolve these problems, we propose a novel DNN that includes a balanced dual image scaler (BDIS) for SR and downscaling. The main contribution of the proposed BDIS is the proposal of an origin referenceable loss (ORL) for downscaling and the balanced training architecture. The proposed ORL is designed to observe the difference between the original and the downscaled images so that the downscaling module directly exploits the information of the original image for its training. However, this ORL can lead to the training imbalance where the downscaling module is relatively overtrained. Therefore, we construct the balanced training architecture by adding the symmetric ORL for SR. The simulation results showed that the proposed BDIS greatly improves the quality of the downscaled images while providing the comparable quality of the super-resolved images compared with the benchmark methods.