Object segmentation based on Gaussian mixture model and conditional random fields

摘要

The feature representation has significantly profound impact on the segmentation accuracy. This paper proposes a feature extract method for conditional random fields (CRF) to segment image into object and background. It divides the pixel into the different semantic model via Gaussian Mixture Model (GMM), then uses the mean values and covariance of every sub-component in GMM for CRF. The proposed method takes the probability distribution of the image as the feature representation for CRF learning. Then it use the CRF to segment the image into semantic parts with the homogeneous appearance based on the minimum energy, which is computed via data costs, label costs and smooth costs. We do performance evaluation experiments on the MSRC-21 benchmark. Experimental results show the proposed method is competitive to the state-of-the-art method for object segmentation.