Face Recognition System Using Codebook Space Information Processing

利用“矢量化码书-空间信息处理算法”开发了一种新的人脸识别技术。具体处理步骤如下：首先将人脸图像分成4X4的小块，通过减去块内的最小强度提取差分强度信息。然后利用理论合成的码本进行矢量量化，通过统计分析每个码本矢量的引用频率提取个人特征信息。最后，将特征信息与数据库进行匹配，进行人物识别。我们开发了有效的滤波程序来消除噪声和不需要的信号成分，并开发了直方图标准化程序来校准人脸的大小。此外，我们引入了“有效识别距离”作为识别度量，以便在更高成功率区域改进识别算法。最后，我们对44个人的220张人脸图像实现了100%的识别成功率。我们将该识别算法应用于面部表情识别。对同一个人的3种面部表情（愤怒、快乐、正常）的识别成功率达到100%。通过评估合适的滤波器尺寸和图像分辨率，发现在真实空间中周期为13mm ~ 14mm或更长的信号分量对面部表情识别非常重要。我们还研究了一种说话人识别技术，以实现高精度的人体识别和身份识别。采用倒谱分析提取特征，矢量量化分类，实现了高性能的说话人识别算法。通过新开发的分层匹配方法和预学习过程，提高了识别速度和成功率。最后，我们对58人290篇不依赖文本的演讲实现了最高97%的识别成功率。少

Novel faco rccognition technology has been developed using "Vector-Quatization Codebook-Space Information Processing Algorithm." Detailed processing steps are as follows. Facial image is first divided into small size blocks (4X4) and differential intensity information is extracted by subtracting minimum intensity within the blocks. Then vector quantization is carried out using theoretically synthesized codebook and personal feature information is extracted by statistically analyzing referred frequenucies of each codebook vector. Finally, matching between the feature information and the database is carried out to identify the person. We have developed effective filtering procedure to eliminate noise and unwanted signal component and histogram standardization procedure to calibrate the size of faces. In addition, we have introduced "effective discrimination distanace" as a recognition measture in order to improve the recognition algorithm at higher success rate region. Finally, we have r … More ealized 100% recognition success rate for 44 person's 220 facial images.We have applied the recognition algorithm into the facial expression recognition. 100% success rate has been realized in recognizing 3 facial expressions (anger, happiness, and normal) of the identical person. It is revealed by evaluating suitable filter size and image resolution that the signal components having a period of 13mm to 14 mm or longer at real space are very important for facial expression recognition.We have also studied a speaker recognition technology for realizing highly accurate human recognition and identification. We have realized high performance speaker recognition algorithm, which utilizes feature extraction by cepstrum analysis and classification by vector quantization. Improvements on recognition speed and success rate were achieved by newly developed hierarchical matching method and pre-learning procedure. Finally, we have realized 97% recognition success rate at maximum for 58 person's 290 text-independent speeches. Less