Adaptive deep learning framework for robust unsupervised underwater image enhancement
Alzayat Saleh, Marcus Sheaves, Dean R. Jerry, Mostafa Rahimi Azghadi
Abstract
One of the main challenges in deep learning-based underwater image enhancement is the limited availability of high-quality training data. Underwater images are often difficult to capture and typically suffer from distortion, color loss, and reduced contrast, complicating the training of supervised deep learning models on large and diverse datasets. This limitation can adversely affect the performance of the model. In this paper, we propose an alternative approach to supervised underwater image enhancement. Specifically, we introduce a novel framework called Uncertainty Distribution Network ( UDnet ), which adapts to uncertainty distribution during its unsupervised reference map (label) generation to produce enhanced output images. UDnet enhances underwater images by adjusting contrast, saturation, and gamma correction. It incorporates a statistically guided multicolour space stretch module (SGMCSS) to generate a reference map, which is utilised by a U-Net-like conditional variational autoencoder module (cVAE) for feature extraction. These features are then processed by a Probabilistic Adaptive Instance Normalisation (PAdaIN) block that encodes the feature uncertainties for the final image enhancement. The SGMCSS module ensures visual consistency with the input image and eliminates the need for manual human annotation. Consequently, UDnet can learn effectively with limited data and achieve state-of-the-art results. We evaluated UDnet on eight publicly available datasets, and the results demonstrate that it achieves competitive performance compared to other state-of-the-art methods in both quantitative and qualitative metrics. Our code is publicly available at https://github.com/alzayats/UDnet . • Underwater image enhancement is challenging but essential for various applications. • Fully-supervised enhancement models need diverse and labelled high-quality images. • Proposed an unsupervised open-source enhancement deep learning model, UDNet. • Trained UDNet on undertwater videos without any labelling and with limited data. • Show that UDNet outperforms 10 state-of-the-art models across 8 public datasets.