基于曲线估计的实时曝光图像增强算法  

Real-Time Exposure Image Enhancement Algorithm Based on Curve Estimation

作　　者：金帅鸿 李国成[1] JIN Shuaihong;LI Guocheng(School of Applied Science,Beijing Information Science and Technology University,Beijing 100192,China)

机构地区：[1]北京信息科技大学理学院,北京100192

出　　处：《昆明理工大学学报（自然科学版）》2024年第1期65-72,共8页Journal of Kunming University of Science and Technology(Natural Science)

基　　金：国家自然科学基金项目(62176073)。

摘　　要：针对曝光错误导致图像质量降低的问题,图像增强旨在不破坏正常曝光区域质量的同时,提高曝光错误区域的质量.然而,在近年来的研究中,使用一个算法解决多种曝光问题并不常见,而且通常需要大量的参数和内存,这不可避免地增加了成本和时间的开销.本文提出了一种新的算法,利用全局-局部感知轻量级Transformer网络和全局-局部光增强曲线,在边缘设备的有限资源下高效的提高图像质量.该轻量级网络主要由全局分支和局部分支两个部分组成.全局分支使用Transformer模块提取最适合的全局参数映射,以区分和调整图像的全局信息.而局部分支获取图像的像素信息,用于估计最佳的局部参数映射.最后,通过迭代运用包含有全局参数映射和局部参数映射的全局-局部光照增强曲线提高了图像质量.在曝光错误数据集上的实验表明:所提出的算法仅需要5%的参数和0.1%浮点运算即可达到与目前STOA算法相当的图像质量,从而显著提高了效率.Addressing the issue of reduced image quality due to exposure errors,image enhancement aims to improve the quality of regions affected by exposure errors without compromising the quality of normally exposed areas.However,in recent research,it is uncommon to find algorithms that address multiple exposure issues simultaneously,and such algorithms often require a significant number of parameters and memory,inevitably increasing costs and time consumption.This paper introduces a novel algorithm that efficiently enhances image quality using a lightweight Transformer network with global-local perception and a global-local luminance enhancement curve,designed to operate effectively within the limited resources of edge devices.The lightweight network consists of two main components:a global branch and a local branch.The global branch utilizes Transformer modules to extract the most suitable global parameter mapping,distinguishing and adjusting the global information of the image.The local branch acquires pixel information from the image to estimate the optimal local parameter mapping.Finally,the iterative application of a global-local luminance enhancement curve containing global and local parameter mappings enhances image quality.Experimental results on an exposure error dataset indicate that the proposed algorithm achieves image quality comparable to current state-of-the-art algorithms with only 5%of the parameters and 01%of the floating-point operations,significantly improving efficiency.

关 键 词：图像增强 光照 深度学习 曝光校正 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]