基于深度学习的水下光学图像超分辨率重建综述

A Review of Underwater Optical Image Super-Resolution Reconstruction Based on Deep Learning

摘要:

图像超分辨率重建旨在从低分辨率图像中生成包含高频细节的高分辨率图像。随着近年来人工智能的快速发展,基于深度学习的超分辨率重建算法取得了突破性进展。然而,水下光学图像通常会产生严重的颜色失真、细节缺失、对比度下降与模糊等多种退化问题,重建难度远高于常规的自然光学图像。目前尚未有文献对基于深度学习的水下光学图像超分辨率重建进行系统性综述。首先,对自然图像退化方式和数据集进行分类总结,结合国内外最新研究现状将基于深度学习的单幅图像超分辨率重建算法分为针对一般退化、已知(非盲)多种退化、未知(盲)多种退化三个方面进行详细总结,为水下应用场景提供参考。然后,介绍了水下光学图像退化方式,归纳了常见的公开数据集,总结并分析了水下光学图像超分辨率重建的最新进展。最后,对该领域未来可能的发展趋势进行了展望。

Abstract:

Image super-resolution reconstruction aims to generate high resolution images containing high-frequency details from low resolution images. With the rapid development of artificial intelligence in recent years, the super-resolution reconstruction algorithm based on deep learning has made a breakthrough. However, underwater optical images usually have many degradation problems, such as serious color distortion, missing details, contrast reduction and blurring, which makes the reconstruction difficulty much higher than that of conventional natural optical images. At present, there is no systematic review of underwater optical image super-resolution reconstruction based on deep learning. First of all, the natural image degradation methods and data sets are classified and summarized. Combining the latest research, the single image super-resolution reconstruction algorithm based on deep learning is summarized from three aspects: general degradation method, non-blind multiple degradation method, and blind multiple degradation method, which provides reference for underwater application scenarios. Then, the degradation mode of underwater optical image is introduced, the common open data sets are summarized, and the latest progress of underwater optical image super-resolution reconstruction is summarized and analyzed. Finally, the possible future development trend of this field is prospected.