【作者】 王斌；

【导师】 何坤；

【作者基本信息】 四川大学 ， 工程硕士（专业学位）， 2021， 硕士

【摘要】 图像前景指的是图像中用户感兴趣对象,是图像最为重要的组成部分,其提取质量关系到内容感知、特征提取以及图像分析等后续任务的完成质量。然而,自然场景图像通常包含了复杂的纹理信息,而图像纹理会弱化前/背景区域颜色的一致性,并在前、背景区域内部形成伪边缘效应,进而造成前景提取效果不佳。此外,由于不同图像一般蕴含不同程度的纹理信息,难以通过简单的平滑预处理有效抑制图像纹理,容易造成图像欠平滑或过平滑。为了弥补传统前景提取算法的不足,本文结合图像视觉效果随着观测距离的变化规律,建立图像多尺度保边平滑模型,结合不同尺度下的图像边缘以及前、背景颜色分布,提出了融合多尺度特征的图像前景提取模型,以实现自然场景图像前景的有效及准确提取。本文主要工作及创新归纳如下:(1)为克服图像纹理对前景提取的负面影响,本文分析了各向同性和异性扩散机理,比较了两者的优缺点,总结了保边扩散核函数的理想条件,提出了新的保边扩散核函数,理论证明了该函数在图像边缘和纹理区域分别执行各向异性扩散和近似各向同性扩散,在平滑图像纹理的同时具有良好的保边性。利用扩散核函数,建立了图像多尺度保边平滑模型,可为后续前景提取过程提供具有不同平滑尺度的图像边缘以及区域颜色分布特征。(2)为从平滑分量中准确提取前景对象,本文根据图像同一区域颜色分布的内聚性以及不同区域间颜色的差异性,选用高斯混合模型描述图像前、背景模式,借助平滑分量的亮度直方图估计区域个数,提升了图像前、背景模式的估计精度。联合图像边缘和前、背景模式,建立了图像前景提取的图表示,提出了图像前景提取的能量泛函,并借助模式参数估计和分割的交替优化过程实现了对平滑分量前景对象的有效提取。(3)为克服固定尺度对前景提取的负面影响,本文研究了图像观察结果与观测距离之间的关系,结合图像多尺度保边平滑与平滑分量的前景提取,建立了融合多尺度特征的前景提取模型。通过交替进行图像平滑与前景提取,可以得到原始图像在不同尺度下的提取结果。为了从合适尺度提取前景,本文分析了相邻尺度上提取结果的相似性,设计了一个迭代终止条件,确保算法终止于恰当尺度,解决了不合适尺度对前景提取的负面影响。实验结果表明,本文提出的保边扩散核函数,能有效平滑图像中的纹理,并保护前景的边缘信息。同时,在BSD500、CMU及GSC数据集的结果显示,本文提出的前景提取算法能够实现对自然图像前景对象的准确及有效提取。更多还原

【Abstract】 Image foreground,the image object of interest,is the most vital portion for arbitrary image,and its extraction quality is related to the completion quality of subsequent tasks such as content perception,feature extraction and image analysis.However,natural scene images usually contain complex texture information,while the image texture will weaken the color consistency of the foreground or background,generate pseudo edges within foreground,and further result in poor foreground extraction.What’s more,since different images generally contain different degrees of texture information,it is difficult to effectively suppress image texture through simple smoothing preprocessing,which easily causes image under or over smoothing.To make up for the shortcomings of traditional foreground extraction algorithms,this thesis constructs an image multiscale edge-preserving smoothing model according to the changing law of image visual effects with the observation distance.Combining image edges and color distributions at different scales,an image foreground extraction model fused with multiscale features is proposed to realize the effective and accurate extraction of image foreground objects.The main work is briefly described as follows:(1)To overcome the negative effect of image texture on foreground extraction,this thesis analyzes the isotropic and anisotropic diffusion mechanisms,compares their advantages and disadvantages,summarizes the ideal conditions of the edge-preserving diffusion kernel function,proposes a new edge-preserving diffusion function,and theoretically proves that this function performs anisotropic diffusion and approximate isotropic diffusion at image edges and textures,respectively,and has good edge preservation while smoothing image textures.Using the diffusion function,an image multiscale edge-preserving smoothing model is established,which can provide image edges and regional color distributions for the subsequent foreground extraction process.(2)To accurately extract foreground objects from smooth components,the Gaussian mixture model is utilized to model the image pattern of foreground and background according to the color cohesion in the same image region and the color difference between different regions.For each smoothing component,the number of image regions is estimated with their intensity histogram,which further improves the estimation accuracy of foreground and background patterns.Combining image edges and foreground pattern,a graph of foreground extraction is constructed,the energy functional of foreground extraction is proposed,and the effective extraction from smoothing component is realized with the aid of alternate optimization of pattern parameter estimation and segmentation.(3)To overcome the negative impact of fixed scale on foreground extraction,this thesis studies the relationship between image observation results and observation distance.Combining image multiscale edge-preserving smoothing and foreground extraction of smooth components,a foreground extraction model fused with multiscale features is proposed.By alternately performing image smoothing and foreground extraction,the extraction results of the original image at different scales can be obtained.To extract the foreground from an appropriate scale,this thesis analyzes the similarity of extracted results on adjacent scales,and designs an iterative termination condition to ensure that the algorithm terminates at an appropriate scale,thereby solving the negative impact of inappropriate scales on foreground extraction.Experiments demonstrate that the edge-preserving diffusion function can not only preserve foreground edges accurately,but also suppress image textures effectively.What’s more,the results on the BSD500,CMU and GSC data sets have proved that the proposed model can achieve accurate and effective extraction for foreground objects in natural images.更多还原