基于四叉树分解和MB_LBP模式的人脸识别算法

Face recognition algorithm based on multi-scale block local binary and quad-tree decomposition

针对多尺度块局部二值模式MB_LBP在分块尺寸选择上的不足,提出了一种基于四叉树分解的MB_LBP算法。算法首先利用图的四叉树分解方法对人脸图像进行高同质性的划分,得到非固定尺寸的图像子块集合;然后对每个图像块利用改进的MB_LBP算法得到图像块的LBP值;最后,将所有的LBP值形成统计直方图得到整幅图像的特征。算法有效考虑了图像块的纹理特性,优化了局部特征值选择,提高了识别正确率。通过在ORL和YALE人脸数据库中进行仿真实验,结果表明,提出的算法相比MB_LBP方法识别正确率提高约10个百分点,并且在针对高分辨率人脸图像时识别速度提高1 s左右。算法适用于高分辨率的人脸图像快速识别。

In order to solve the block size selection weakness of multi-scale block local binary pattern MB_LBP, an improved MB _LBP algorithm based on quad -tree decomposition is proposed. At first, the presented algorithm utilized quad-tree decomposi- tion method of graph to divide the image with high homogeneity, and got a set of image sub-blocks with non-fixed size. Then, an improved MB_LBP algorithm was used to compute each image sub-block＇ s value of LBP. At last, the whole image＇ s feature val- ue was obtained by constructing the statistical histogram of all LBP values. Because the presented algorithm efficiently considers the texture characteristics of image blocks and optimized the choose of local feature values, it improves the recognition accuracy. Through the simulation experiment in face database on ORL and YALE, the experimental results show that the presented algorithm has higher recognition efficiency than MBLBP method, the correct recognition rate is increased by about 10% , and the recogni- tion speed is increased by about ls when apply in the field of high resolution face image recognition. The presented algorithm is suitable for fast recognition of face images with high resolution.