Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (20): 237-244.DOI: 10.3778/j.issn.1002-8331.2206-0132

• Graphics and Image Processing • Previous Articles     Next Articles

Attention Exposure Fusion Network for Low-Illumination Image Enhancement

BAO Yifeng, YANG Degang   

  1. College of Computer and Information Sciences, Chongqing Normal University, Chongqing 401331, China
  • Online:2023-10-15 Published:2023-10-15

面向低照度图像增强的注意力曝光融合网络

包易峰,杨德刚   

  1. 重庆师范大学 计算机与信息科学学院,重庆 401331

Abstract: Low-illumination image enhancement is used to improve the brightness of low-illumination images and fully exploit the hidden information in the images. Recently, low-light image enhancement methods based on deep learning have achieved significant advantages over traditional methods. In order to solve the problem that existing methods may lead to overexposure of bright details or loss of dark details in images, and cannot take into account both dark and bright details in images. A multi-scale attention exposure fusion network is proposed. First, a set of images with different exposure magnifications are generated from a single image. Then, each image is passed through the proposed multiscale attention block to improve the low-illumination image enhancement performance. In particular, the proposed residual context block models the global context of the input features to obtain rich global features. The selective kernel block reselects and calibrates multiple features to enhance useful information while suppressing useless information to improve the processing power of the network. Meanwhile, the fusion block constructs an accurately exposed, detail-rich image by fusing well-exposed regions in images with different exposure magnifications. The PSNR and SSIM performance are improved by 0.72 dB and 0.08 respectively compared to the SID method when compared with the mainstream methods in recent years on the SID dataset. Experimental results show that the improved method has better detail reproduction and color reproduction than the existing methods.

Key words: attention mechanisms, image enhancement, low-illumination images, exposure fusion

摘要: 低照度图像增强用于提高低照度图像的亮度并充分挖掘图像中的隐藏信息。最近,基于深度学习的低照度图像增强方法相对于传统方法取得了显著的优势。为了解决现有方法可能会导致图像亮部细节过曝或暗部细节丢失、无法兼顾图像暗部和亮部细节的问题。提出一种多尺度注意力曝光融合网络。从单个图像中生成一组具有不同曝光倍率的图像。每张图像通过提出的多尺度注意力模块提高低照度图像增强性能。提出的残差上下文模块对输入特征进行全局上下文建模,获取丰富的全局特征。选择性内核模块利用注意力机制结合具有不同感受野的特征,同时保留了它们独特的互补特征。融合模块通过融合不同曝光倍率图像中曝光良好的区域来构建一张曝光准确、细节丰富的图像。在SID数据集上与近年来主流方法进行了对比,PSNR和SSIM性能相比SID方法分别提升了0.72?dB和0.08。实验结果表明,改进的方法比现有方法拥有更好的细节还原和色彩还原能力。

关键词: 注意力机制, 图像增强, 低照度图像, 曝光融合