MOESR: MULTI-OBJECTIVE EVOLUTIONARY ALGORITHM FOR IMAGE SUPER-RESOLUTION

Senqiao Yang, Guodong DU, Zhang Bohao, Ho-Kin Tang, Sim Kuan Goh

Primary Area: representation learning for computer vision, audio, language, and other modalities

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.

Keywords: MULTI-OBJECTIVE EVOLUTIONARY ; SUPER-RESOLUTION

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: In recent years, deep neural networks have gained substantial traction in the field of super-resolution. However, existing deep learning methods primarily focus on enhancing the peak signal-to-noise ratio~(PSNR) of images, resulting in suboptimal performance across various evaluation metrics and a lack of fine details in image visual quality. To address these limitations, we introduce a comprehensive algorithmic framework, Multi-Objective Evolutionary Algorithm for Image Super-Resolution (MOESR), which aims to achieve a balanced optimization of multi-objective in image super-resolution. Specifically, MOESR first decomposes the multi-objective super-resolution problem into sub-problems and employs a novel approach to generate an initial population for the evolutionary algorithm. Subsequently, it enhances mutation, crossover, and update processes using an improved differential evolution algorithm, yielding a more Pareto-efficient set of solutions. Compared to traditional gradient-based methods, our approach does not require gradient calculations for each objective. As a result, it avoids issues such as gradient vanishing or local optima. Furthermore, our method has lower computational complexity, making it particularly advantageous for addressing high-dimensional problems and deep networks. Extensive experiments are conducted on five widely-used benchmarks and two multi-objective tasks, resulting in promising performance compared to previous state-of-the-art methods. In addition, our approach can not only address multi-objective optimization problems but also represents the first method capable of addressing the balance between objective and perceptual metrics. Our code will be released soon.

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: 691