Segmenting hair regions from human images facilitates many tasks like hair analysis and hair style trends forecast. However, hair segmentation is quite challenging due to large within-class pattern diversity and between-class confusion resulted from complex illumination and similar appearance. To solve these problems to some extent, this paper proposes a novel coarse-to-fine hair segmentation method. Firstly, the recently published "active segmentation with fixation (ASF)" is used to coarsely define a candidate region with high-recall (but possibly low-precision) of hair pixels and exclude considerable part of the backgrounds which are easily confused with hair. Then the graph cuts (GC) method is applied to the candidate regions to perform more precise segmentation, by incorporating image-specific hair information. Specifically, Bayesian method is employed to select some reliable hair and background regions (seeds) among the ones over-segmented by mean shift. SVM classifier is then learnt online from these seeds and explored to predict hair/background likelihood probability, which is used as an initialization for performing GC algorithm. Experimental results demonstrate the approach outperforms existing hair segmentation methods. To validate the generality, the paper extends the method and achieves good results on the public databases of horse, car and aeroplane classes.