Contrastive Semi-Supervised Learning for Underwater Image Restoration via Reliable Bank

Despite the remarkable achievement of recent underwater image restoration techniques, the lack of labeled data has become a major hurdle for further progress. In this work, we propose a mean-teacher based Semi-supervised Underwater Image Restoration (Semi-UIR) framework to incorporate the unlabeled data into network training. However, the naive mean-teacher method suffers from two main problems: (1) The consistency loss used in training might become ineffective when the teacher's prediction is wrong. (2) Using L1 distance may cause the network to overfit wrong labels, resulting in confirmation bias. To address the above problems, we first introduce a reliable bank to store the “best-ever” outputs as pseudo ground truth. To assess the quality of outputs, we conduct an empirical analysis based on the monotonicity property to select the most trustworthy NR-IQA method. Besides, in view of the confirmation bias problem, we incorporate contrastive regularization to prevent the overfitting on wrong labels. Experimental results on both full-reference and non-reference underwater benchmarks demonstrate that our algorithm has obvious improvement over SOTA methods quantitatively and qualitatively. Code has been released at https://github.com/Huang-ShiRui/Semi-UIR.

1. Introduction

Due to light refraction, absorption and scattering in underwater scenes, images taken in the water usually suffer severely from color distortion, low contrast and blur. Images with these defects tend to be less visually appealing and can potentially hinder the well-functioning of underwater robotic systems. Recently, many deep learning based methods [5-7,24,51] have been proposed to address image restoration problems. Numerous efforts have also been devoted to the specific domain of underwater image restoration [11], [17], [20], [22], [49]. Compared with traditional methods that mostly rely on hand-crafted priors, deep learning based solutions are able to deliver superior restoration results due to their data-driven nature. Figure 1.

Examples from different benchmarks. (a) Shows real-world underwater images from uieb [22] with degraded images (first and second row). (b) Shows the uwcnn training set [21] (synthesized based on the image formation model) and (c) shows the euvp dataset [17] (synthesized by gan). The ambient light and color cast of (b) and (c) are quite different from that of (a).