Unsupervised Multiple Representation Disentanglement Framework for Improved Underwater Visual Perception

Visual perception is a key component for many underwater applications. However, visual imaging data typically suffer from different types of distortion, which seriously affect their security and reliability. Underwater image enhancement (UIE) is a crucial step to improve the perceptual quality of the environment. Recent deep learning techniques for UIE are predominantly supervised in nature and, hence, require imaging data of the same scene to be acquired in distortion-corrupted and distortion-free environments, which is impractical, especially when a sufficient sample size is needed for reliable network training. As a remedy, we propose an unsupervised multiple representation disentanglement (UMRD) framework for improved underwater visual perception. UMRD employs a unified multidomain unsupervised cycle-consistent adversarial translation to simultaneously disentangle the domain label indicating image type, content features reflecting structural semantics, and style information reflecting appearance from all-domain underwater images. The disentangled content feature and target style information are then utilized to generate the target distortion-free image. Furthermore, we present the underwater multidistortion image enhancement data set, a large-scale collection of 19 K underwater images (of poor and good quality) that can be used for improved underwater visual perception (e.g., decolorization, deblurring, and denoising). Experimental results demonstrate that UMRD, trained using unpaired data, exhibits prominent capabilities of image quality enhancement and performance improvement for subsequent processing tasks, i.e., object detection and semantic segmentation.