Résumé: Low-light image enhancement technology is widely used extensively in industry, but it is also a challenging aspect of computer vision tasks. Although existing methods are committed to adjusting the overall brightness of the image, they ignore the problems of region excessive exposure and uneven illumination enhancement. They also introduce new problems, such as noise generation and tone distortion, during the process. To solve these problems, we propose a two-step deep learning-based method that combines a proposed attention-guided network with hierarchical global priors (GPANet) with a Mapping Curve function. In the proposed method, GPANet is first used as a deep curve estimation network to generate a self-adaptive enhanced mapping feature map. Subsequently, a pixel-wise curve is utilized to aid in the generation of natural tone-enhanced images. To evaluate the performance of the proposed method, we conducted comparative experiments with different methods on commonly used public datasets, and the proposed method achieved superior results in terms of naturalness image quality evaluator and lightness order error.

Langue: Anglais