Color Image Segmentation Combining Rough Depth Information
A novel color image segmentation method is presented in this paper. Firstly a Luv color histogram based method is used to estimate the color bandwidth, then a mean shift algorithm with adaptive color bandwidth is employed to pre-segment the input image. N
- PDF / 1,077,769 Bytes
- 10 Pages / 439.37 x 666.142 pts Page_size
- 12 Downloads / 215 Views
2
University of Science and Technology of China, Hefei, Anhui, China [email protected] Institute of Intelligent Machines, Chinese Academy of Sciences, Hefei, Anhui, China
Abstract. A novel color image segmentation method is presented in this paper. Firstly a Luv color histogram based method is used to estimate the color bandwidth, then a mean shift algorithm with adaptive color bandwidth is employed to pre-segment the input image. Next, a boundary detection algorithm based machine learning is used to calculate the probability boundary of objects from both depth and color information. Then, a correction procedure is performed by mapping the depth boundary onto the color image. Finally, Graph cut is used to segment color image based on Gaussian Mixture Model which is built with the above pre-segmentation and correction results. The experimental results show that the segmentation algorithm is an effective one. It can effectively segment an image into some semantic objects. Keywords: Image segmentation · Depth map Gaussian mixture model · Graph cut
1
· Adaptive mean shift ·
Introduction
The use of depth information has been the subject of a number of vision-related tasks such as image segmentation and retrieval. For the task of image segmentation, depth information has obvious advantages as compared with color information. Firstly it is invariant to lighting and/or texture variation; secondly it is invariant to camera pose and perspective change. Therefore using depth can potentially enable successful segmentation independent of illumination or view, significantly expanding the range of operation conditions. In this background, the color image segmentation combining depth information has received much attention. Recently, many efforts have been made. Crabb et al. employed depth map to extract foreground objects in real time[1]. In their segmentation the color image is only used to support a small fraction (1%-2%) of the pixels which are not solved by the depth threshold. Another interesting use of depth is implemented in [2]. Several channels of the depth camera are used and combined with the color channel to assist background subtraction. A new indoor scene dataset, completing with accurate depth maps and dense label coverage, is introduced in c Springer-Verlag Berlin Heidelberg 2015 H. Zha et al. (Eds.): CCCV 2015, Part I, CCIS 546, pp. 448–457, 2015. DOI: 10.1007/978-3-662-48558-3 45
Color Image Segmentation Combining Rough Depth Information
449
[3] by Nathan Silberman et al. Their model which is evaluated on this dataset is inclined to solve the indoor image segmentation. Meir Johnathan Dahan et al. present a technique [4] which is the most relevant work with ours. However, their results depend on the color image segmentation method greatly. In this paper we segment image with an improved mean shift algorithm with adaptive color bandwidth. Then we strengthen segmentation results combining color and depth effectively. This algorithm can weaken the illumination changes and object shadow effects on image segmen
Data Loading...
