基于特征融合的人脸素描自动生成系统研究

RESEARCH ON FEATURE FUSION-BASED AUTOMATIC FACIAL SKETCH GENERATION SYSTEM

人脸素描作为一门独特的艺术,以其逼真的手法表现人物特征和个性,广泛应用于法律实施、数字娱乐和生活工艺等方面。鉴于其广阔的应用前景,人脸素描自动生成已成为计算机视觉领域的热门话题之一。针对当前人脸素描使用单一图像特征,素描效果不够精细,建立的模型对光照变化比较敏感等问题,提出在使用马尔可夫随机场(MRF)模型和置信度传播算法(BP)基础上进行光照鲁棒性细节技术设计,融合多种主流视觉特征进行人脸素描自动生成。首先,将人脸素描分割成交叠图像块,使用Do G、Dense SIFT、Gray,LBP,gist等特征定义能量函数,建立MRF模型;然后,利用BP算法进行推理,估计测试照片块对应的素描块;最后,对素描块交叠区域进行处理,拼接成一幅完整的人脸素描。实验结果表明,使用Sum-Product置信度传播算法进行推理,对素描块重叠区域用平均法进行拼接,素描合成达到较好视觉效果,且具有较好光照鲁棒性。

As a unique art,facial sketch expresses people’s feature and character with its realistic skill,and is widely used in law enforcement,digital entertainment,life technology and so on.In view of its broad application prospects,automatic facial sketch generation has become one of the hot topics in computer vision field.Aiming at the problems that current facial sketch uses single image feature,the sketch effect is not fine enough,and the built model is also very sensitive to illumination changes,etc.,in this article we present to conduct the technical design for robustness details of illumination based on using MRF model and belief propagation （BP）algorithm,and to carry out automatic facial sketch generation by combining a couple of mainstream visual features.First,the facial sketch is divided into overlapping image patches,and the features definition energy functions of DoG,Dense SIFT,Gray,LBP gist,etc.,are used to build Markov random field （MRF）model.Then,BP algorithm is used for inference in order to estimate the corresponding sketch patch of the testing photo patch. Finally,the overlapping regions of the sketch patches are processed and assembled into a complete facial sketch.Experimental result illustrates that the sketch synthesis effect achieves better effect by using Sum-Product BP algorithm for inference and using average method to assemble overlapping regions of sketch patches,and has better illumination robustness as well.