基于通道协同与色彩校正的RAW数据到RGB图像重建方法  

A Method for RAW Data to RGB Image Reconstruction Based on Channel Synergism and Color Correction

作　　者：刘昱 陈永祺 王瀚 刘人赫 汪少初[2] Liu Yu;Chen Yongqi;Wang Han;Liu Renhe;Wang Shaochu(School of Microelectronics,Tianjin University,Tianjin 300072,China;Tianjin Institute of Surveying and Mapping Co.,Ltd.,Tianjin 300381,China)

机构地区：[1]天津大学微电子学院,天津300072 [2]天津市测绘院有限公司,天津300381

出　　处：《天津大学学报(自然科学与工程技术版)》2025年第4期395-405,共11页Journal of Tianjin University:Science and Technology

基　　金：天津市自然科学基金资助项目(22JCZDJC00270);天津市研究生科研创新资助项目(2021YJSO2B12).

摘　　要：图像传感器直接输出的原始RAW数据需经过重建操作后才能转换为RGB图像,用于重建RGB的传统多步骤图像信号处理(ISP)算法往往具有流程复杂、灵活性低、误差累积等问题,而现有的一些基于深度学习的图像信号处理算法存在未能充分利用RAW图像模式信息与未能充分利用RGB通道相关性等局限.基于此,本文设计了基于深度学习的ISP-Net模型,针对上述不足进行改进.该模型包含模式信息复原、跨通道重建和色调调整3个主要模块.首先,模式信息复原模块通过掩码提取RAW图像的颜色通道信息,并通过通道间的融合操作最大化地学习RAW图像的模式信息;其次,跨通道重建模块通过通道间相互指导的方式,充分利用通道相关性提升图像重建质量;最后,色调调整模块通过构建色调调整矩阵,提升颜色复原的准确性.所设计的模型在合成数据集与真实数据集上均进行了测试.在主观评价上,本文方法在色彩准确度、纹理细节重建方面更接近真实图像;在客观指标上,特别地,在噪声强度为15时,本文方法的峰值信噪比在Kodak、McMaster、WED-CDM和Urban 100数据集上均为最优指标,分别为32.33 dB、32.44 dB、32.22 dB及31.67 dB,最大可优于同数据集次优指标0.24 dB.在真实数据集上的峰值信噪比为28.98 dB,相较于传统图像信号处理方法FlexISP、SEM和ADMM,峰值信噪比均高出至少17 dB,相比于基于深度学习的RAW图像重建方法Deepjoint和MGCC-B峰值信噪比分别提升了4.03 dB和1.55 dB.The original RAW data output by image sensors needs to be reconstructed into RGB images.Traditional multistep image signal processing(ISP)algorithms often face challenges such as complex processes,low flexibility,and error accumulation.Furthermore,some deep-learning-based ISP algorithms struggle to fully utilize the mode information of RAW images and the correlation between RGB channels.To address these deficiencies,ISPNet—a deep learning model designed for improved RGB image reconstruction—is proposed herein.The model comprises a mode information restoration module,a cross-channel reconstruction module,and a tonal adjustment module.The mode information restoration module extracts color channel data from RAW images using masks and learns mode information through channel fusion.The cross-channel reconstruction module utilizes channel correlations to enhance image reconstruction quality through mutual guidance between channels.Finally,the tonal adjustment module employs a tonal adjustment matrix for precise color restoration.Experiments conducted on synthetic and real datasets reveal that in terms of subjective evaluation,the ISP-Net model achieves good color accuracy and texture detail,aligning closely with real images.In terms of objective metrics,particularly with noise intensities of 15,the peak signal-to-noise ratio(PSNR)of ISP-Net exhibits optimal performance across the Kodak,McMaster,WEDCDM,and Urban 100 datasets,with values of 32.33 dB,32.44 dB,32.22 dB,and 31.67 dB,respectively.These values outperform the second-best metrics within the same datasets by up to 0.24 dB.On real datasets,the PSNR reaches 28.98 dB,which surpasses the traditional ISP algorithms,namely FlexISP,SEM,and ADMM,by at least 17 dB.Compared to those of the deep-learning-based RAW image reconstruction algorithms such as Deepjoint and MGCC-B,the PSNR is improved by 4.03 dB and 1.55 dB,respectively.

关 键 词：图像信号处理 RAW图像重建 模式信息复原 跨通道重建 色调调整 

分 类 号：TN911.73[电子电信—通信与信息系统]