Low-Light Image Enhancement Based on Mutual Guidance Between Enhancing Strength and Image Appearance

PATTERN RECOGNITION AND COMPUTER VISION, PRCV 2023, PT XI(2024)

引用 0|浏览18
暂无评分
摘要
The existing low light image enhancement (LLIE) methods primarily aim at adjusting the overall brightness of the image, which are prone to produce the over-enhancement issue, such as over-exposure and edge halo. Therefore, it is desirable to improve the visibility of originally dark regions of an image, while preserving the naturalness of the originally bright regions. Based on this motivation, we propose a simple but effective mutual guidance module, which builds a mutual guidance process between a pixel-wise enhancing strength map and an edge-aware lightness map. Based on this module, the image appearance information such as illumination and structure can be effectively propagated onto the enhancing strength map. By integrating this module into the ZeroDCE++ model, the over-enhancement issue like over-exposure and edge halo can be greatly alleviated. We have conducted extensive experiments to validate the effectiveness and the superiority of our model. Compared with many state-of-the-art unsupervised and supervised LLIE methods, our model achieves a much better visual effect as it consistently keeps the naturalness during the enhancement process. Our model also has better or comparable performance than its counterparts in quantitative comparison with various image quality assessment metrics.
更多
查看译文
关键词
Low-light image enhancement,Mutual guidance module,Unsupervised model,Over-enhancement
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要