基于散集图的图像前景区域提取算法研究

发布时间：2018-11-04 20:58

【摘要】：图像前景区域提取是指从一幅图像中提取出感兴趣的任意形状的前景区域。前景区域提取是图像处理中的基本操作,也是视觉特效制作中最重要和最常用的操作,在杂志、二维图形艺术、电视、广告等领域也有广泛的应用。前景提取算法主要可以根据认为图像中是否含有半透明像素分为两类：图像分割和前景抠图。本文简单介绍了图像前景提取的现有主流算法,并对多种算法的优缺点进行了分析与比较。目前主流的前景提取算法几乎都是需要大量的人工交互过程。我们可以发现摄影作品在形式上区别于人的正常视觉感受,在于它利用相机镜头的光学特性制造了一个兴趣中心,有了中心的画面,当然会引起人的注意。摄影爱好者利用镜头与眼睛的结合,把眼睛的感受和镜头的表现天衣无缝地曝光在底片上,将自己感兴趣的区域传递给观看者。现在的单反相机或无反相机等拍摄器材在推销时,经常都会强调产品可以制造更佳的散景表现。散景,简单地说就是指主体与背景之间那种前清后蒙的效果。而这种效果突出了主体的立体感,强化了主体,一定程度上区分了前景背景的区域。就算是普通相机拍摄的图片,也会有轻微的散景效果。散景程度在图像处理中我们可以用图像的散焦图来表示。本文先对图像的散焦进行放大处理,使得普通图片也具有跟专业相机拍摄效果一样的散焦图,再将散焦图与源图像再融合起来,得到类似专业相机拍摄出来的效果。本文又从摄影者的角度出发,重新定义了前景区域的具体概念-即拍摄者所想要突出的区域和镜头最终的表现出的显著区域。本文用散焦图和显著图来表示这两个特性。基于这一想法,本文提出了一种不需要进行人工交互的基于散焦图的前景提取算法。本文对图像数据先进行图像分割,再将图像经过放大的散焦图与图像的显著图结合起来作为图像分块的特征,并训练了一个基于svm的分类器将前景从背景中分离出来。在测试图像中,如果图像分段在已训练完的测试模型中取得高可信度,就将该分段标志为前景。本文提出的算法在MSRA显著物体检测图像集和Flickr图像集上均进行了实验,通过实验结果验证了本文提出的前景定义的合理性和算法的可行性。
[Abstract]:Image foreground region extraction refers to extracting any shape foreground region of interest from an image. Foreground region extraction is not only the basic operation in image processing, but also the most important and commonly used operation in visual special effects production. It is also widely used in magazines, two-dimensional graphic art, television, advertising and other fields. The foreground extraction algorithm can be divided into two categories according to whether the image contains translucent pixels: image segmentation and foreground matting. This paper briefly introduces the existing mainstream algorithms of image foreground extraction, and analyzes and compares the advantages and disadvantages of various algorithms. At present, the mainstream foreground extraction algorithms almost require a large number of manual interaction processes. We can find that photography is different from normal visual feeling in form, because it makes use of the optical properties of camera lens to create a center of interest, with a central picture, of course, it will attract people's attention. Photographers use the combination of the lens and the eyes to seamlessly expose the eye's feelings and the lens's performance to the negative and transmit the area of interest to the viewer. Current cameras such as SLR or non-reflective cameras are often marketed with emphasis on products that produce better performance. Scattered scenery, in short, refers to the effect of the former Qing Dynasty and the later Mongolia between the subject and the background. This effect highlights the stereoscopic sense of the subject, strengthens the subject and, to some extent, distinguishes the region of foreground background. Even pictures taken by ordinary cameras have a slight scatter effect. In image processing, the degree of scatter can be expressed by the image's astigmatic image. In this paper, the defocus of the image is amplified first, so that the common image has the same effect as the professional camera, and then the defocus image and the source image are fused together again to get the effect of the similar professional camera. From the perspective of the photographer, this paper redefines the specific concept of foreground region, that is, the region that the photographer wants to highlight and the region in which the lens finally displays. In this paper, the defocus diagram and the salient graph are used to represent these two properties. Based on this idea, a defocus map based foreground extraction algorithm is proposed. In this paper, the image data is segmented first, then the enlarged defocus image is combined with the salient image as the feature of the image block, and a classifier based on svm is trained to separate the foreground from the background. In a test image, the segmentation is marked as a foreground if it is highly reliable in the trained test model. The proposed algorithm is tested on both MSRA salient object detection image set and Flickr image set. The experimental results verify the rationality of the proposed foreground definition and the feasibility of the algorithm.
【学位授予单位】：复旦大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41

【相似文献】