High-Fidelity Retrieval of Low - Light Underwater Images using Advanced Deep Learning Model for Enhanced Visual Quality

Resumen

The improvement of the visual quality of the image of the underwater objects in the low-light conditions is a must in such applications as marine monitoring, self-driving cars, and detection of the objects underwater. Such pictures are often characterized by color variations, reduced brightness and high amount of noise because of the rapid extinction of the light in water. In order to deal with these difficulties, we examine two state-of-the-art deep learning models: Zero-DCE++, a zero reference curve estimation model which can enhance the low-light images without using ground truth supervision and U-former, a transformer based architecture that is targeting the high fidelity image restoration. The two models are applied at preprocessing stages in order to enhance contrast, color fidelity and general visibility before sending the enhanced images to detection or segmentation process. Blind/Reference-less Image Spatial Quality Evaluator (BRISQUE), Natural Image Quality Evaluator (NIQE), Lightness Order Error (LOE) and Colourfulness Index (CI) are quantitative measures of an improved performance. The experimental outcomes show that there is significant enhancement of the visual quality and strength of IQE - 3.90, RISQUE - 28.40, LOE- 740, CI - 52.10 over conventional convolution-based methods.