【作者】 张翱； 朱宇； 孙瑾秋； 张艳宁；

【Author】 ZHANG Ao;ZHU Yu;SUN Jinqiu;ZHANG Yanning;School of Computer Science,NorthwesternPolytechnical University;School of Astronautics,NorthwesternPolytechnical University;

【机构】 西北工业大学计算机学院； 西北工业大学航天学院；

【摘要】 成像过程中大多数相机的传感器能接收到的辐照度范围远小于真实场景,此外,各种噪声的存在以及相机和被摄景物之间可能出现的相对运动,将导致拍摄的图片具有较窄的动态范围,并且存在一定程度的运动模糊。传统的基于去模糊模型对图像进行复原的过程,多是迭代的对模糊核进行估计,再利用估计的模糊核对图像进行非盲卷积复原。已有的模糊核估计方法通过一系列先验信息来计算模糊核,如空间分布里的重尾高斯先验,L0、L1正则化等,但它们都是基于全局或基于图像块的,需要对图像中每一个像素的值都进行计算,算法工作量大,而且这些方法未考虑拍摄图像在曝光时间过长或过短时出现的高亮极暗区域,导致模糊核估计不准确,复原结果中亦会产生振铃现象。为了解决这一问题,我们对不同曝光下的多组模糊图片进行分析,针对图像中存在的过饱和与欠饱和区域,定性定量的分析其对模糊核估计的负作用。根据分析得到的结果,可预先去除不利于模糊核估计的部分,以提高估计结果的精度并复原出具有更多细节,振铃更少的清晰图像。相机相应函数(CRF)曲线建立了由CCD传感器阵列接收到的场景照度到低动态范围(LDR)图像像素比特深度之间的非线性映射关系。我们首先对真实辐照度图像与LDR图像的动态范围进行分析,将两类CRF曲线(DoRF数据集中提供的真实曲线以及由MATLAB仿真的曲线)作用于一组不同曝光时间下的辐照度图像,通过ToneMap显示辐照度图像,观察其与LDR图像的区别。对图像进行对比可得在曝光时间较短的情况下,LDR图像的暗弱区域相比辐照度图像丢失了更多的细节,在曝光时间较长的情况下,LD R图像将会出现较多过饱和区域,而辐照度图像则会显示出更多的细节,动态范围更广,含有极少的高亮极暗区域。其次,去模糊模型是由经过CRF曲线量化后的图像与模糊核卷积而成,而实际模糊图像形成过程为辐照度图像与模糊核相作用,加上曝光时间对CCD的影响,再经过CRF曲线映射而得,我们将辐照度图像用不同的CRF曲线分别以两种不同的模糊方式来做处理,针对像素级差值进行对比分析。结果表明同一CRF作用下,高亮极暗区域由于已经达到饱和状态,不同模糊方式得到的图像间差值最小,图像梯度边缘部分由于像素值跨度较大,CRF曲线对其进行量化时会出现较大误差,故差值最大,而像素值跨度大的区域包含的信息更多,也更有利于模糊核的估计。本文的分析结果可为图像去模糊作支撑,通过剔除存在的高亮极暗区域,并选取更利于估计模糊核的部分,减少模糊核估计的误差,能有效的抑制复原结果中产生的振铃现象。更多还原

【Abstract】 Most camera sensors can receive less irradiance than that of the real scene in the image processing,in addition,the existence of various noises and the relative motion between camera and the object affect the image quality.These factors lead to a narrow dynamic range of images and have motion blur in most cases.Traditional process of image restoration based on deblurring estimates blur kernel of image first,and then recover the sharp image using non-blind convolution methods with the kernel iteratively.Existing methods rely on using a series of image prior to gain the blur kernel,such as heavy-tailed Gauss priori and the L0/L1 regularization on the characteristics of the spatial distribution.But they are based on the global image or image-block which have to compute the data of every single dimension,besides,there is no consideration of either over-saturated or under-saturated areas resulting from the long/short exposure time.It causes the inaccuracy of blur kernel estimation with some noise around it,and the recovered image will also introduce a lot ringing artifacts near strong edges.To address the problem,we analyze the negative effects of these regions on blur kernel estimation qualitatively and quantitatively analyze.According to the analysis results,the parts that are not conducive to estimation can be removed in advance to improve the accuracy of the results and restore a sharp image with more details and fewer ringing artifacts.The Camera Response Function(CRF) curve establishes the non-linear mapping relationship between the scene illumination received by CCD sensor array and the pixel bit depth of Low Dynamic Range(LDR) image.Firstrly,we analyze the dynamic range of real irradiance images and LDR images.Two kinds of CRF curves(real curves provided by DoRF data set and curves simulated by MATLAB) are applied to a group of irradiance images with different exposure times.The irradiance images are displayed through Tonemap to observe the differences between them and LDR images.By comparing the two kinds of images,we can get that under the condition of short exposure time,the information of scene will sink in the dark area of LDR image.In the case of long exposure time,more over-saturated areas will appear in the LDR image,while the irradiance image shows more details with a wider dynamic range,and contains very few bright or dark areas.Secondly,thedeblurring model is formed by the convolution of images quantified by CRF curve with blur kernel,while the actual forming process of blur image is the interaction between irradiance image and the blur kernel within the impact of exposure time on CCD array,and finally obtained by mapping CRF curves.We use different CRF curves to process the irradiance image in two different kinds of blur methods which mentioned above,and conduct comparative analysis of the difference on pixel-level.The results show that under the action of the same CRF,the difference between the bright and dark regions is the smallest due to they have reached saturation state already.On the contrary,the edge of image gradient is the largest because of the large span of pixel value,so there will be a large error in quantifying with CRF curves.And the regions with large spans of pixel values contains more information,also is more advantageous to the estimation of blur kernel.The analysis results in this paper can support for imagedeblurring.By eliminating the bright and dark areas,and selecting parts that are more conducive to estimating the blur kernel,it can reduce the error of estimation,and the ringing artifacts generated in the restoration results can be suppressed effectively.更多还原