基于子块LBP和最优相关的素描人脸合成

The Face Sketch Synthesis Based on Sub-blocks LBP and Optimization

提出一种素描人脸合成算法,其主要实现的功能是给出一幅光学脸,合成一幅素描脸。为了合成素描图片,对人脸区域进行分块,利用欧氏距离从训练集中提取与待合成目标相近的粗选块系列;使用子块切分的LBP纹理筛选对粗选块系列进行再提取,得到几个与待合成目标更加相近的精选块系列;提出基于最优相关的逐次定位法,即确定首行首块,依次计算相邻块间的相关系数,求得最优块,最终合成一副完整的素描人脸。通过对待合成目标进行性别过滤,以进一步提升合成精度。经实验验证,该算法在FERET数据库下多个训练集,测试集80幅人脸的情况下,合成精度达到92%左右,验证了素描人脸合成算法准确性和有效性。

A sketch face synthesis algorithm is proposed,and the main function is to give an optical face and a sketch face will be synthesized.To synthesize sketch images,the face region is divided into patches for learning,and the Euclidean distance is used to extract rough blocks which are similar with synthetic target from the training set.The LBP texture filter combined with sub-blocks was presented to extract further in order to obtain several more reasonable selected blocks.Successive locating method based on optimal correlation is put forward.It determines first line and first block,and the adjacent blocks correlation coefficient are calculated to obtain corresponding optimal block,and eventually form a complete sketch faces.The article considered that gender characteristics can be used to further improve the accuracy of synthesis and recognition rate.By testing80faces from the FERET database,the experimental verification shows that recognition rate reaches92%and it also verify the accuracy and validity of the sketch face synthesis algorithm in multiple training set.