加权合成的嵌入式隐Markov模型人脸识别

Weighted synthesis embedded-hidden Markov model for face recognition

嵌入式隐 Markov模型能提取人脸的二维主要特征并对姿态和环境的变化具有较好鲁棒性 .讨论了嵌入式隐 Markov模型的进一步改进及其实现 .首先分析了形成观察向量的采样窗大小和其二维 DCT系数项数的不同对人脸识别结果的影响 ,然后确定最优的采样窗大小和其二维 DCT系数项数 .鉴于不同角度的照片包含信息量的不同 ,提出了一种加权合成的模型参数重估算法 .重估模型参数时 ,首先计算每幅脸像相对应的模型参数 ,然后进行加权合并 ,权值由迭代公式求得 ,训练结束后用一个合成的模型来表示一个对象 .采用基于该方法的原型系统对 ORL人脸库进行测试 ,识别正确率达到了 99.5 % .

Embedded Hidden Markov Model (E-HMM) can extract the main features of faces and has necessary robustness property in treating diversities of poses and lighting environments. This paper deals with the performance improvement of E-HMM and its implementation. First, The effects of the size of the sampling window and the terms of 2D-DCT coefficients of every image block on the face recognition accuracy are analyzed, and the optimal sizes of the sampling window and the terms of 2D-DCT coefficients of every image block are selected based on the analysis results. In view of the fact that the contribution entropies of different photos to the final face E-HMM are different, a new weighted synthesis method for re-estimating E-HMM parameters is developed and described in detail. During the period of re-estimating the E-HMM parameters, every training sample was represented by one E-HMM, the model parameters for every sample were obtained firstly, then the different model parameters were synthesized to one model through weighted method, and the weights were adaptively calculated in the training stage. After the training, one person's face's images were represented by one E-HMM parameters. The running experimental results with ORL (Olivetti Research Ltd.) face database show that the recognition rate with this new approach has reached about 99.5%.