基于自适应截断模拟曝光和无监督融合的低照度真彩色图像增强算法  

Low-light True Color Image Enhancement Algorithm Based on Adaptive Truncation Simulation Exposure and Unsupervised Fusion

作　　者：韩永成 张闻文[1] 何伟基[1] 陈钱[1] HAN Yongcheng;ZHANG Wenwen;HE Weiji;CHEN Qian(School of Electronic and Optical Engineering,University of Science and Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学电子工程与光电技术学院,南京210094

出　　处：《光子学报》2023年第9期237-251,共15页Acta Photonica Sinica

基　　金：国家自然科学基金(No.61875088),江苏高校“青蓝工程”项目。

摘　　要：针对图像增强算法中普遍存在暗部区域细节丢失、亮部区域过度增强等问题,提出了基于自适应截断模拟曝光和深度融合的增强算法。对原始低照度图像进行模拟曝光,通过卷积网络学习曝光序列对应的权重图并在网络内部实现加权融合,获得增强结果。在生成模拟曝光序列的过程中,对图像进行亮暗区域分割,然后对其进行截断性自适应伽马校正,最后通过引导滤波降噪获得合适的曝光序列。获得多曝光序列后,通过基于空洞卷积的上下文聚合网络,实现快速灵活地加权融合,得到最终增强结果。收集了大量公开数据集,并使用微光夜视相机和三通道真彩色相机收集了实验室环境测试集,在不同数据集上和经典主流算法进行了对比实验。实验结果表明,本算法的NIQE、PSNR和SSIM指标都是最好的,其中NIQE降低了4.49%,PSNR提高了4.28%,SSIM提高了1.94%。此外,算法的色彩还原效果也很好,色差指标是所有算法中最小的,在8.71×10^(-2)lx照度下,本算法色差减小了14.83%,在1.02×10^(-2)lx下减小了3.05%。本文算法可以明显提高图像亮度和对比度,鲁棒性较好,不会产生过度增强现象,有效恢复图像细节的同时兼顾色彩信息,增强结果真实自然。Low-light true color image enhancement is an important branch in image processing.The images obtained in low-illuminance environments often have low brightness,low contrast,noise,and color distortion.Due to the complexity and diversity of target scenes and imaging equipment,it is difficult to directly obtain satisfactory high-quality images in low-illumination environments.There are many problems in the information content of these low-light true color images,and the image is not good for viewing and is not conducive to advanced image tasks in the later stage.Aim at solving the problems of detail loss in the dark area and excessive enhancement in the bright area,a low-light true color image enhancement algorithm based on adaptive truncation simulation exposure and deep fusion is proposed.The algorithm has the function of brightness suppression.We design an adaptive truncation simulation exposure method and use an unsupervised network model to fuse the exposure sequence to achieve a flexible and efficient fusion of multiple exposure images of fixed size.First,an exposure sequence about the original low-light true color image is generated by simulation,and then the convolutional network is used to learn the weight map corresponding to the exposure sequence.We can obtain the final enhanced results by weighted fusion within the network.Most classic simulated exposure algorithms either map the image linearly or use existing enhancement algorithms such as histogram equalization.The number of simulated exposures is often determined artificially in pursuit of as many exposure sequences as possible covering different brightness levels,which results in many redundant images in the simulated exposure sequence.To address these problems,we propose an adaptive gamma correction method which can effectively avoid this redundancy.The light and dark regions of the image are segmented first,then truncated adaptive gamma correction is carried out.Finally,the appropriate exposure sequence is obtained by guided filter denoising.After

关 键 词：彩色夜视 图像增强 自适应曝光 曝光融合 卷积神经网络 图像质量评价 

分 类 号：TN223[电子电信—物理电子学]