Abstract: A single-image automatic hair detection algorithm for the application of automatic generation of facial caricature is proposed.First, hair regions in training images are labeled manually.Hair location probabilistic masks and hair color probabilistic distribution functions are estimated from these labels.Second, an energy function of the test image is built according to the estimated prior distributions of hair location and hair color likelihood.The energy function is further optimized by graph cut algorithms and initial hair region is obtained.Finally, K-means algorithm and image post-processing techniques are applied in the initial hair region to precisely segment the final hair region.Experimental results show that the computational cost of the proposed method is low.The average processing time is about 300 ms per image.The average hair region detection accuracy is above 90%.Our proposed method satisfies the requirements of automatic generation of facial caricature.