PROJECT TITLE :

An Underwater Image Enhancement Benchmark Dataset and Beyond

ABSTRACT:

Due to its importance in marine engineering and aquatic robotics, underwater image enhancement has gotten a lot of attention. There have been a slew of algorithms for improving underwater images in the last few years. Instead than using real-world photos, these algorithms are typically tested on artificial datasets. As a result, it's hard to say how well these algorithms will perform on real-world photographs or how we can measure success in the field. Using large-scale real-world photos, we conduct the first-ever complete perceptual study and analysis of underwater image augmentation. 950 real-world underwater photographs, 890 of which have comparable reference images, make up the Underwater Image Enhancement Benchmark (UIEB) in this work. As difficult data, we use the remaining 60 underwater photographs for which we were unable to obtain appropriate reference images. Using this data, we've been able to undertake a qualitative and quantitative analysis of the most current underwater picture enhancement methods. The proposed UIEB for training Convolutional Neural Networks can be generalised by using an underwater image enhancement network (named Water-Net) trained on this benchmark as a baseline (CNNs). By comparing current algorithms to benchmarks and developing Water-Net, we can better understand where the field of underwater picture enhancement is headed in the future. This secure URL provides access to the data and the source code.