基于RAW域的低光照图像质量增强方法  

Low-Light Image Enhancement Based on RAW Domain Image

作　　者：陈龙[1] 张一驰 吕张凯 丁丹丹[1] Chen Long;Zhang Yichi;Lyu Zhangkai;Ding Dandan(School of Information Science and Technology,Hangzhou Normal University,Hangzhou 311121)

机构地区：[1]杭州师范大学信息科学与技术学院,杭州311121

出　　处：《计算机辅助设计与图形学学报》2023年第2期303-311,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(62171174);浙江省自然科学基金(LY20F010013);国家级大学生创新创业训练计划(202110346065).

摘　　要：考虑原始图像(RAW图像)中的噪声未经过相机ISP的复杂非线性映射,更加容易建模,提出基于RAW域对低光照图像进行质量增强,以获得清晰的、高质量的图像.首先,将由相机传感器捕获的RAW图像进行线性插值,获得RGGB共4个通道的彩色图像,并模拟不同曝光程度的图像;其次,设计神经网络模型来学习不同曝光程度的RAW图像与正常曝光图像的映射关系,该神经网络采用了自编码器结构,并嵌入了通道注意力模块,以提取并增强图像中的纹理与细节特征;最后,在神经网络训练时,引入了结构相似性损失和梯度下降损失,引导网络生成与参考图像之间结构相似性和色彩关联性较高的高质量图像.该方法在SID数据集上进行了训练和测试,PSNR达到了29.7380 dB,MPSNR达到了30.2334 dB,均高于目前流行方法(如EnlightenGAN,Zero-DCE,SID,Residual和ALEN等);在主观质量方面,该方法生成的图像质量显著优于比较方法,无明显噪声和色斑伪影,颜色也更加真实细腻.It is easier to model the noise in RAW image data(RAW images)than in RGB images because RAW images provide the original data without the nonlinear mapping of camera ISP.Therefore,this paper proposes a low-light image quality enhancement method in the RAW domain for clear and high-quality im-ages.Firstly,the RAW image captured by the camera sensor is linearly interpolated to obtain a four-channel RGGB color image.Secondly,we generate more RAW images by applying different exposure levels on this RGGB image.Finally,we develop a neural network to learn the mapping between the RAW images with different exposure levels and reference images.Our proposed neural network adopts the autoencoder struc-ture where the channel attention module is incorporated to extract and enhance the latent features of images.In the training,we design a new loss function that combines both structural similarity loss and gradient de-scent loss to guide the network to produce high quality images with high structural similarity and color relevance to the reference image.The proposed method is trained and tested on the See-In-the-Dark(SID)dataset.It achieves average 29.7380 dB PSNR and 30.2334 dB MPSNR,which outperforms state-of-the-art methods EnlightenGAN,Zero-DCE,SID,Residual,and ALEN.In terms of subjective quality,the images enhanced by the proposed method have no obvious noise and color spot artifacts,looking more visually pleasing and appealing than the images from previous methods.

关 键 词：低光照图像 图像增强 卷积神经网络 通道注意力机制 拜耳模式图像 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]