Semi-Supervised Color Image Segmentation Research And Application

On the basic computional module of semi-supervised hidden Markov random fields (HMRFs) with constraints, a modified color image segmentation method is proposed in this paper. MeanShift algorithm is applied to get supervision information, cluster number and initial values for cluster centroids. In HMRFs, no method is mentioned to obtain cluster labels, cluster number and centroids. These initialization issues can be solved via MeanShift algorithm. Some modifications are made about K-Means algorithm. The third part with cannot-link constraint is removed. Besides, the update of distance parameters is also removed due to divergence.Texture is taken into the feature vector. Histogram is obtained via Gabor filter, and combined with initial feature.The distance is calculated separately. The experimental result is very encouraging. Color images can be segmented effectively with the new method.Finally, the new segmentation method is applied to face detection to demonstrate its application value. For simple pictures, skin color can be used to detect face. Pixels in each clusters are analyzed according to color information. In certain color range, the detection accuracy is better than the one of detecting pixels sequently.