Using Vision Transformers (ViT) For Low-Level Vision Enhancement

Abstract

Restoring and enhancing underwater images is a significant issue in image processing and computer vision. Poor underwater imaging quality is caused by the scattering and absorption of light by underwater contaminants. Images taken underwater frequently suffer from quality issues, such as low contrast, poor sight (due to the absorption of natural light), blurred details, changing colors, additive noise, blurred effects, and uneven illumination, etc. The study of underwater image analysis has gained a lot of attention and achieved substantial advancements during the past few decades. The current techniques can broaden the application of underwater photography while improving image contrast and resolution. Traditional image enhancement techniques have some drawbacks when applied directly to underwater optical environments; hence, some specific algorithms, such as histogram-based, retinex-based, and picture fusion-based algorithms, are proposed. Deep learning has recently shown a strong potential for creating results that are satisfying and have the right colors and details, but these methods significantly increase the size of the image processing inference models and therefore cannot be applied or deployed directly to the edge devices. Recently, Vision Transformers (ViT)-based architectures are producing incredible results. In recent years, there has been more interest in transformers. Their interactions between image content and attention weights can be thought of as a convolution that changes in space, and their self-attention mechanism is good at simulating long-distance dependencies and global features. The suggested approach is a pipeline based on a context-aware lightweight vision transformer with the goal of improving image quality without sacrificing the naturalness of the image, as well as reducing the inference time and size of the model. In this study, we trained a deep network-based transformer model on two standard datasets, i.e., Large-Scale Underwater Image (LSUI) and Underwater Image Enhancement Benchmark Dataset (UIEB), so that the network becomes more generalized, which subsequently improved the performance. Our real-time underwater image enhancement system shows superior results on edge devices. Also, we provide a comparison with other transformer-based methods. Overall findings indicate that the suggested method has produced underwater images of higher quality than the original input underwater images, which had a high noise ratio and more color disruption.