LOW-LIGHT IMAGE ENHANCEMENT USING A CONVOLUTIONAL ATTENTION MECHANISM IN THEZERODCE FRAMEWORK

Abstract

Image exposure correction and low-light image enhancement are important prob lems in computer vision due to their significant impact on image quality and down stream vision applications. Existing enhancement approaches often suffer from is sues such as noise amplification, color distortion, loss of structural details, and limited adaptability to varying illumination conditions. Although recent deep learning-based methods achieve improved enhancement quality, many of them rely on computation ally expensive architectures and large-scale paired datasets for training. The current thesis provides an in-depth discussion of methods used for exposure correction and introduces two efficient zero-reference enhancement models based on the ZeroDCE framework. While the former is extended using the CBAM for more precise spatially adaptive feature refinement, the latter one utilizes a new technique called the Exposure Guided Attention (EGA). Both models rely on the self-supervised training using the losses without the need for pair-supervision during the process. Experimental results show that the use of the introduced architectures leads to improvements in terms of image quality, preservation of the image structure and perceptual similarity while be ing computationally-efficient. In particular, the newly-proposed framework based on the utilization of EGA achieves significant improvements compared to the baseline Ze roDCE method as it increases the value of PSNR by 6.85 dB (from 10.57 dB to 17.42 dB) and the SSIM score by 0.249 (from 0.233 to 0.482).