Detail Loss in Super-Resolution Models Based on the Laplacian Pyramid and Repeated Upscaling-Downscaling Structure
Keywords: Super-Resoltuion, Laplacian Pyramid-based detail loss, Repeated Upscaling-Downscaling Process, Supervised learning
TL;DR: We introduce the Laplacain Pyramid-based detail loss with a Repeated Upscaling and Downscaling Process to enhance the high-frequency components of super-resolution images.
Abstract: With advances in artificial intelligence, image processing has also gained significant interest. Image super-resolution, in particular, is a vital technology closely related to real-life applications, as it enhances the quality of existing images. Since enhancing details is important in the super-resolution task, it is often necessary to activate pixels that appear only at high frequencies, distinct from low frequencies.
In this paper, we propose a method that generates a detail image separately from the super-resolution image. This approach introduces a loss function designed to enhance detail, allowing the model to generate an upscaled image and a detail image independently, with control over each component. Consequently, the model can focus more effectively on high-frequency data, resulting in an improved super-resolution image. Our loss function utilizes detail images based on the Laplacian Pyramid, which is widely used in image reconstruction. The multi-level property of the Laplacian Pyramid is well-suited for applying upscaling and downscaling repeatedly.
Our experiments demonstrate that a structure applying the repetition of upscaling and downscaling integrates effectively with our detail loss control. The results show that this structure efficiently extracts diverse information, enabling the generation of improved super-resolution images from multiple low-resolution features. We conduct two types of experiments. First, we construct a simple CNN-based model incorporating the Laplacian Pyramid-based detail control and a repeated upscaling and downscaling structure. This model achieves a state-of-the-art PSNR value of 38.48 dB, surpassing all currently available CNN-based models and even some attention-based models without additional special techniques. Second, we apply our methods to existing attention-based models on a small scale. In all the experiments, attention-based models using our detail loss show improvements compared to the original models. These experiments demonstrate that our detail control loss effectively enhances performance, regardless of the model's structure in the super-resolution task.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13157
Loading