【摘要】 为了可靠地测量颅骨与面皮之间的相关性,提高颅骨识别能力,本文提出了一种基于最小二乘正则相关性分析的颅骨识别方法。构建颅骨和面皮的统计形状模型,将高维颅骨和面皮映射到低维形状参数空间;基于最小二乘正则相关性分析提取颅骨和面皮的主相关信息,构建整体相关性分析模型,用于度量颅骨与面皮的整体相关性;然后,考虑到颅骨不同区域的相关度不同,将颅骨分为前额、眼睛、鼻子、嘴巴和轮廓5个区域,基于最小二乘正则相关性分析提取颅骨和面皮的主相关信息,分区域构建区域相关性分析模型,用于度量颅骨与面皮之间的局部细节相关性;最后,分别采用整体相关性分析模型和区域相关性分析模型度量颅骨与面皮的匹配关系,计算颅骨与面皮库中每个面皮之间的匹配分数,具有最高匹配分数的面皮即为正确的识别结果,从而实现颅骨身份识别。实验结果表明,整体相关性分析模型的识别准确率为85.2%,单一区域相关性分析模型中轮廓区域的识别准确率最高,鼻子区域的识别准确率最低;通过融合5个区域建立的相关性分析模型,识别准确率高达95.2%,基于区域融合的方法优于整体相关性分析方法。更多还原

【Abstract】 To measure the correlation between the skull and the face reliably and improve the skull recognition ability,a skull identification method based on least squares canonical dependency analysis(LSCDA)was proposed. First,a statistical shape model of the skull and facial skin was constructed,and the high-dimensional skull and facial skin were mapped to the low-dimensional shape parameter space. Second,the main relevant information about the skull and facial skin was extracted based on LSCDA,and an overall correlation analysis model was constructed to measure the relationship between the skull and the skin. By considering the difference in the correlations between different regions of the skull,the skull was divided into five regions:the forehead,eyes,nose,mouth,and contour. Based on LSCDA,the main relevant information about the skull and facial skin was extracted and constructed by region. A regional correlation analysis model was used to measure the local detail correlation between the skull and the face. Finally,a global correlation analysis model and a regional correlation analysis model were used to measure the matching relationship between the skull and facial skin,and the matching score between each face in the skull and facial skin database was calculated. The face with the highest matching score yielded the correct recognition result for achieving skull identification. The experimental results reveal that the recognition accuracy of the overall correlation analysis model is 85. 2%. The recognition accuracy of the contour region in the single region correlation analysis model is the highest,whereas that of the nose region is the lowest. The correlation analysis established by fusing the five regional models indicates that the recognition accuracy rate is as high as 95. 2% and that the method based on regional fusion is better than the overall correlation analysis method.更多还原