Saliency Detection of Image and Video and a Proposed Approach using Superpixel-Level Markov Random Field Model

• Main Authors: Wen-Wen Chang, 張雯雯
• Other Authors: 貝蘇章
• Format: Others
• Language: en_US
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/54649695203001439495

碩士 === 國立臺灣大學 === 電信工程學研究所 === 102 === Saliency, also known as visual attention, refers to the areas distinct from its surroundings that human observer would focus at a glance. Saliency detection benefits many computer vision tasks, and extensive efforts have been devoted to achieving better saliency detection performance. We observe that most of the previous works are hard to deal with the non-homogeneous color distribution within an object. Motivated by this observation, we consider the spatial structure between image regions to obtain better results. In this thesis, a proposed approach for image saliency detection and its extension for video saliency detection are introduced. The approach is based on background prior and superpixel-level Markov Random Field (MRF) model. First, we separate the image into middle-level superpixels and extract low-level features (color, texture energy, and defocus level) within each superpixel. Then, we build up a Markov-Random-Field (MRF) on the superpixels and adopt simplified propagation technique to optimize the superpixel saliency. Afterward, we refine this superpixel-level solution to pixel-level saliency map. Experimental results demonstrate that our proposed method is promising as compared to the state-of-the-art methods in two public available datasets.