【作者】 杨振；

【导师】 李迎松；

【作者基本信息】 哈尔滨工程大学 ， 电子科学与技术， 2023， 硕士

【摘要】 多曝光图像融合(Multi-exposure image fusion,MEF)是获得高动态范围图像(High dynamic range,HDR)的有效方式之一。MEF方法致力于解决图像细节信息丢失、局部亮度色彩失真、融合光晕伪影及多场景适应性差等问题。此外,现有的MEF方法大多需要具有小曝光差异的长序列原始图像才能获得较好的融合结果,即输入多曝光图像堆栈曝光依次由欠、中以及过曝光图像组成,而当原始图像较少且具有较大曝光差异时融、合图像质量将显著下降。导向滤波作为一种局部线性引导式滤波器,自问世以来便在MEF任务中展现了极大的活力,不论是针对MEF任务中细化基本权重图还是对图像分层进行处理都有较为广泛的应用。针对上述MEF算法存在的融合问题,本文基于导向滤波方法的主要工作如下:第一,针对多尺度结构块分解的多曝光融合方法(MSPD-MEF)存在细节损失的问题,且目前引入边缘保留因子的方法(MESPD-MEF)优化过程繁琐,正则化常数随着多尺度分解的过程也存在优化问题。因此,提出一种基于导向滤波保边的多尺度SPDMEF方法(GFMSPD-MEF)。首先将保边的导向滤波纳入MSPD-MEF,将其退化为一个具有边缘保持滤波器代替均值滤波隐式进行结构块分解。此外,还设计了一种更为灵活的自适应权值进一步保留亮区和暗区的细节,并且所提方法也可以拓展到多尺度框架中,既适用于静态场景,也适用于动态场景,相较于MESPD-MEF仅仅产生了略高的计算时间,生成的融合图像细节信息较为丰富,有较好的边缘保持效果。第二,针对双尺度图像融合采用传统的均值滤波后的图像融合存在光晕伪影以及多场景适应性的问题,提出一种基于导向滤波图像分层的多曝光融合算法。其中,利用图像分层及多尺度金字塔融合的方法来解决融合光晕伪影的问题,在基层和细节层两个维度联合多种设计的权值来解决不同场景图像、增强图像细节特征信息以及保持整体亮度色彩的能力。所提方法能够很好的适应不同场景下的融合任务,且融合图像在整体亮度色彩保持、全局细节信息恢复以及抵抗融合光晕方面都有优异的表现。第三,针对原始图像较少且具有较大曝光差异时融合图像质量将显著下降,且基于导向滤波多尺度分解的图像融合方法(MGFF)中所提的互补性权值不利于极端曝光下获取的问题,为了充分获取极端曝光图像的细节信息,利用导向滤波进行图像多尺度分解,提出一种基于导向滤波的多尺度两曝光图像融合算法。通过导向滤波将图像分解为一个基层和多个细节层,然后设计了分别逐级的曝光权重和全局梯度权重分配策略来挖掘包含在基层和细节层中的图像信息以重建图像。所提方法在极端曝光的条件下生成的融合图像颜色鲜艳、细节清晰、表现自然,符合人类视觉感知习惯。更多还原

【Abstract】 Multi-exposure image fusion(Multi-exposure image fusion,MEF)is one of the effective ways to obtain high dynamic range images(High dynamic range,HDR).The MEF method is committed to solving the problems of loss of image detail information,local brightness and color distortion,fusion halo artifacts,and poor adaptability to multiple scenes.In addition,most of the existing MEF methods require a long sequence of original images with small exposure differences to obtain good fusion results,that is,the input multi-exposure image stack exposures are sequentially composed of under-exposure,medium-exposure,and over-exposure images.While the number if original image is small and there is a large exposure difference,fusion image quality will be significantly reduced.Guided filtering,as a local linear guided filter,has shown great vitality in MEF tasks since its inception.Whether it is for refining the basic weight map or processing image layers in MEF tasks,it has a wide range of applications.In view of the fusion problem existing in the above MEF algorithm,the main work of this paper based on the guided filtering method is as follows:First,for the problem of detail loss in the multi-scale structural patch decomposition multiexposure fusion method(MSPD-MEF),which is an advanced fusion quality with the fastest running time,a multi-exposure fusion with edge-preserving structural patch decomposition(MESPD-MEF)was proposed.While the optimization process of the edge preservation factors introduced is cumbersome and the calculation is large,the regularization constant introduced also has optimization problems with the process of multi-scale decomposition.Therefore,a guided filtering edge-preserving multi-scale SPD-MEF method(GFMSPD-MEF)is proposed.First,the edge-preserving guided filter is incorporated into MSPD-MEF,and it is degenerated into an edge-preserving filter instead of mean filter.In addition,a more flexible adaptive weight is designed to further preserve the details between bright and dark areas,and the proposed method can also be extended to a multi-scale framework,which is suitable for both static and dynamic scenes.Compared with MESPD-MEF,only a slightly higher calculation time is generated,and the generated fusion image has richer detail information and better edge preservation effect.Secondly,aiming at the problem that halo artifacts exist in dual-scale image fusion using traditional mean filtering and image features of different scales need to be guided to filter different filtering radii and blur coefficients,a dual-scale feature enhancement MEF algorithm combined with image layering is proposed.Among them,the method of image decomposition and multi-scale pyramid fusion is used to solve the problem of halo artifact fusion.The weights of various designs are combined in the two dimensions of the basic level and the detail layer to adept different scene images,enhance the image detail feature information and maintain the overall brightness color.Third,the quality of the fused image will decrease significantly when there are few original images with large exposure differences,and the complementary weights proposed in the image fusion method(MGFF)are not conducive to the acquisition under extreme exposure.In order to fully obtain the detailed information of extreme exposure images,guided filtering is used to decompose images at multi-scale,and a multi-scale two-exposure image fusion algorithm based on guided filtering is proposed.The image is decomposed into a base layer and multiple detail layers by guided filtering,and then a step-by-step exposure weight and global gradient weight assignment strategy is designed to mine the image information contained in the base layer and detail layer to reconstruct the image.The fusion image generated by the proposed method under extreme exposure conditions has bright colors,clear details,and natural performance,which is in line with human visual perception habits.更多还原