An Application of MAP-MRF to Change Detection in Image Sequence Based on Mean Field Theory
- PDF / 2,911,052 Bytes
- 13 Pages / 600 x 792 pts Page_size
- 94 Downloads / 194 Views
An Application of MAP-MRF to Change Detection in Image Sequence Based on Mean Field Theory Qiang Liu Laboratory for Computational Neuroscience, Departments of Neurological Surgery and Electrical Engineering, University of Pittsburgh, Pittsburgh, PA 15213, USA Email: [email protected]
Robert J. Sclabassi Laboratory for Computational Neuroscience, Departments of Neurological Surgery and Electrical Engineering, University of Pittsburgh, Pittsburgh, PA 15213, USA Email: [email protected]
Ching-Chung Li Laboratory for Computational Neuroscience, Departments of Neurological Surgery and Electrical Engineering, University of Pittsburgh, Pittsburgh, PA 15213, USA Email: [email protected]
Mingui Sun Laboratory for Computational Neuroscience, Departments of Neurological Surgery and Electrical Engineering, University of Pittsburgh, Pittsburgh, PA 15213, USA Email: [email protected] Received 24 December 2003; Revised 22 October 2004 Change detection is one of the most important problems in video segmentation. In conventional methods, predetermined thresholds are utilized to test the variation between frames. Although certain reasonings about the thresholds are provided, appropriate determination of these parameters is still problematic. We present a new approach to change detection from an optimization point of view. We model the video frames and the change detection map (CDM) as Markov random fields (MRFs), and formulate change detection into a problem of seeking the optimal configuration of the CDM. Under the MRF assumption, the optimal solution, in the sense of maximum a posteriori (MAP), is obtained by minimizing the energy function associated with the MRF which is designed by utilizing the prior knowledge of noise and contextual constraints on the video frames. An algorithm that computes the potentials and optimizes the solution is constructed by applying the mean field theory (MFT). The experimental results show that the new method detects changes accurately and is robust to noise. Keywords and phrases: change detection, image processing, Markov random field, mean field theory, video segmentation.
1. INTRODUCTION Content-based video processing has been widely studied and is supported by a number of standards, such as MPEG-4 for video object-based compression and MPEG-7 for video content description [1, 2]. These standards involve functionalities that rely on segmenting video sequences into semantic regions or video objects. Change detection, which generates an initial segmentation mask, usually constitutes the first step of video segmentation [3, 4]. Much research effort has been devoted to change detection in recent years [5, 6, 7, 8]. Most existing approaches focus on thresholding which contains two essential steps of defining a metric function of intensity variation and choos-
ing a proper threshold to be applied to the metric function. The key issue of these methods is to determine the threshold. However, it is often problematic choosing the threshold, since a large threshold removes noise as well as
Data Loading...
