Underwater Color Restoration Using U-Net Denoising Autoencoder

Abstract

Visual inspection of underwater structures by vehicles, e.g. remotely operated vehicles (ROVs), plays an important role in scientific, military, and commercial sectors. However, the automatic extraction of information using software tools is hindered by the characteristics of water which degrade the quality of captured videos. As a contribution for restoring the color of underwater images, Underwater Denoising Autoencoder (UDAE) model is developed using a denoising autoencoder with U-Net architecture. The proposed network takes into consideration the accuracy and the computation cost to enable realtime implementation on underwater visual tasks using end-toend autoencoder network. Underwater vehicles perception is improved by reconstructing captured frames; hence obtaining better performance in underwater tasks. Related learning methods use generative adversarial networks (GANs) to generate color corrected underwater images, and to our knowledge this paper is the first to deal with a single autoencoder capable of producing same or better results. Moreover, image pairs are constructed for training the proposed network, where it is hard to obtain such dataset from underwater scenery. At the end, the proposed model is compared to a state-of-the-art method.

An Extensive Literature Review on Underwater Image Colour Correction

“…The authors in [ 47 ] created the UW Denoising Autoencoder (UDAE) model, a deep learning network built on a single denoising autoencoder employing U-Net as a Convolutional Neural Network (CNN) architecture, as a contribution to restoring the colour of UW images. The suggested network considers accuracy and computation cost for real-time implementation on UW visual tasks utilising an end-to-end autoencoder network, resulting in improved UW photography and video content.…”

“… One example presented by Hashisho et al ( a ) is the input image, ( b ) the result produced by UGAN and ( c ) the result produced by UDAE. Images from [ 47 ]. …”

Cast-Gan: Learning To Remove Colour Cast From Underwater Images

“…We validate the proposed loss function with an ablation study and compare Cast-GAN with four state-of-the-art methods: two neural networks, namely U-Net Denoising (U-Net) [8] and Underwater Scene Prior Inspired Image Enhancement (UWCN) [12], and two physics-based methods, namely Depth-dependent Background Light (DBL) [16] and Underwater Haze Line (UWHL) [19]. UWHL and Cast-GAN are the only methods explicitly aiming to remove colour cast.…”

“…Table 1 shows that, when compared against the original image, Cast-GAN is selected by the participants 60.2% of the times. UWHL is Original DBL [16] U-NET [8] UWCN [12] UWHL [19] Cast-GAN Note how Cast-GAN removes the cast and enhances details, while maintaining realistic colours, especially in images with heavy casts. Since U-Net can only process images of specific resolutions, we select here the image provided by authors, whose resolution is the closest to the original image's aspect ratio.…”

“…Underwater image enhancement can use neural networks [8][9][10][11][12][13][14][15] or physics-based models [9,[16][17][18][19]. Neural networks for underwater image enhancement are trained using images taken closer to the water surface, which however are still affected by some colour cast [8]. Alternatively, training images are synthesised with neural networks [14,15] or from reference images with physics-based models [9][10][11][12].…”