Domain Adversarial Learning Towards Underwater Image Enhancement

Underwater images are degraded due to the absorption and scattering of light inside the water. The underwater degradation causes the loss of information in terms of texture, style, color, and minute detail of edges and hence the degraded images are not useful in many higher-order applications. Several deep learning techniques are explored by researchers across the globe for the same. Further, deep learning networks learn the distribution of degraded training samples and fail when there is a deviation due to a change in water type. This paper proposes an encoder-decoder network that preserves the image content, texture, and style while maintaining overall global similarity by capturing the inherent distribution of the training samples. To overcome the deviation due to a change in water type, a classifier network is induced in the latent space of encoder-decoder architecture. The classifier loss and adversarial loss in the classifier network ensure the learning across domains and avoid setting priors on captured distribution. Hence, the proposed model is robust against the change of water type and can be deployed in real-life without retraining. To train the model, we use attenuation coefficients of underwater environments at different depths to recreate 5430 paired underwater images from the Underwater Image Enhancement Benchmark (UIEB) dataset with six distinct types of water. Our proposed model enhanced the degraded images among different degradation levels due to depth and water type variations. The proposed model is evaluated on UIEB and EUVP benchmark databases. The performance of the proposed model is verified against twenty-two state-of-the-art methods in terms of underwater reference and no-reference image quality assessment metrics.