结合空间注意力机制与多尺度分辨率融合的低光照图像增强方法  

A low-light image enhancement method combining spatial attention mechanism and multi-scale resolution fusion

作　　者：王秀君 钱蓉蓉 任文平[1] WANG Xiujun;QIAN Rongrong;Ren Wenping(School of Information Science and Engineering,Yunnan University,Kunming 650504,China)

机构地区：[1]云南大学信息学院,昆明650504

出　　处：《激光杂志》2023年第6期86-91,共6页Laser Journal

基　　金：国家自然科学基金(No.61701433);云南省科技计划项目(No.2018FB099)。

摘　　要：为了解决低光照条件下拍摄造成的对比度过低、颜色失真等多种图像内容退化问题,提出了一种结合空间注意力机制与多尺度分辨率融合的图像增强方法。首先基于Retinex理论,利用带有空间注意力机制的分解网络将低光照图像分解为光照图与反射图,并且在光照图中,采用伽玛校正对光照图拉伸光照对比度。然后在融合网络中,将校正过的光照图与反射图通过U-net网络,使高、低分辨率的特征图充分融合来生成最终的色泽度饱满、细节信息丰富的结果。提出的算法在LOL-dataset数据集的结构相似度指标为0.8032,比同类最优算法高出6.37%。In order to solve various image content degradation problems such as low contrast and color distortion caused by shooting in low light conditions,this paper proposes an image enhancement method that combines spatial attention mechanism and multi-scale resolution fusion.First,based on the Retinex theory,a decomposition network with a spatial attention mechanism is used to decompose the low-light image into a luster map and a refraction map,and in the luster map,gamma correction is used to stretch the light contrast of the luster map.Then in the fusion network,the corrected luster map and refraction map are passed through the U-net network to fully fuse the high-and low-resolution feature maps to generate the final result with full color and rich details.The results of the proposed algorithm on the LOL-dataset show that the structural similarity index is 0.8032,which is 6.37%higher than the best algorithm in the same class.

关 键 词：RETINEX理论 空间注意力机制 图像增强 伽玛校正 U-net网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]