基于图像处理的肤色聚类分割人脸检测算法研究

发布时间：2018-03-17 16:05

本文选题：人脸检测　切入点：粒子群算法　出处：《新疆大学》2017年硕士论文　论文类型：学位论文

【摘要】：人脸富含独特的生物特征信息,具有唯一性和高可辨识性,使其作为身份认证已广泛应用到智能监控、社区安防、金融支付等生活重要领域。人脸检测作为人脸图像信息处理的第一步,能否准确、快速的检测并标定出一张图片中人脸的位置、数量有着重要的意义和研究价值。科技的进步致使信息安全不断受到新技术的挑战,人们逐渐从传统的认证方式转向数字化安全性更高的生物特征认证方式,如指纹、虹膜、人脸等生物特征。其中人脸具有稳定性、唯一性,安全性高,难以被复制,使其作为生物特征识别技术之一受到越来越多研究者的关注近年来。随着模式识别、人工智能等领域的快速发展人脸检测的研究也获得了很大的进步,目前针对正常环境下的清晰正面人脸已实现准确快速的检测。但在现实生活中,由于光照、遮挡、复杂背景等条件的影响导致拍摄到的人脸图像并不是清晰的正面照,这会给人脸图像后续的处理带来一定的困扰。针对目前人脸检测算法存在的一些不足之处,为了提升检测技术的准确率和鲁棒性,本文在深入研究了大量传统的人脸检测算法后,提出了相关的优化改进算法。研究内容主要有如下:第一、针对在强光、遮挡等复杂环境下的人脸图像,本文提出一种基于改进粒子群(Particle Swarm Optimization,PSO)与K均值聚类肤色分割的人脸检测方法。该方法首先将待测图片变换到YCgCr彩色空间,利用肤色信息在YCgCr彩色空间分布的集中性,在该色彩空间使用改进的粒子群与K均值聚类综合的方法进行肤色分割;为有效的去除人脸区域以外噪声,得出候选人脸区,需要对分割后得到的肤色区域使用二值形态学和人脸几何形状特征处理;最后通过改进的AdaBoost算法对候选人脸区域进行验证。经实验结果显示,使用本文方法进行人脸检测正确率较高,同时算法的鲁棒性和适应性好。第二、针对受光照不均影响的人脸图像,本文提出一种基于肤色分割和特征定位的多人脸检测方法。该方法首先在RGB空间检测图像是否有色彩偏差,若存在色彩偏差则采用改进的参考白算法进行光照补偿;接着将处理后的图片转换到YCbCr颜色空间进行肤色分割,并通过改进的AdaBoost算法进行人脸检测得到候选区域;然后采用经过大量人眼训练实验得到的先验知识在候选人脸区域标记出人眼;最后输出带人眼定位的人脸图像。
[Abstract]:Face is rich in unique biometric information, unique and highly identifiable, so it has been widely used in intelligent monitoring, community security, as identity authentication, As the first step of face image information processing, whether face detection can accurately, quickly detect and calibrate the position of face in a picture, With the development of science and technology, information security is constantly challenged by new technologies. People are gradually changing from traditional authentication to biometric authentication with higher digital security, such as fingerprint, iris, etc. As one of the biometric recognition techniques, human face has attracted more and more attention in recent years. With the development of pattern recognition, it is difficult to be copied because of its stability, uniqueness, security and so on. In recent years, as one of the biometric recognition techniques, face has attracted more and more attention. The rapid development of face detection in artificial intelligence and other fields has also made great progress. At present, the clear frontal face detection in normal environment has been realized accurately and quickly. But in real life, due to illumination, occlusion, Due to the influence of complex background and other conditions, the face image taken is not clear positive image, which will bring some troubles to the subsequent processing of human face image. There are some shortcomings in the current face detection algorithm. In order to improve the accuracy and robustness of the detection technology, after deeply studying a large number of traditional face detection algorithms, this paper proposes a related optimization improvement algorithm. The main contents of the research are as follows: first, in order to improve the accuracy and robustness of the detection technology, In this paper, a face detection method based on improved particle swarm optimization (PSO) and K-means clustering is proposed for face image detection in complex environments, such as occlusion, etc. The method first transforms the image under test into YCgCr color space. Using the color information in YCgCr color space distribution centrality, the improved particle swarm and K-means clustering method is used to segment the skin color in this color space, in order to effectively remove the noise outside the face region, the candidate face area is obtained. It is necessary to use binary morphology and facial geometry feature processing to segment the skin color region. Finally, the improved AdaBoost algorithm is used to verify the candidate face region. The experimental results show that, The accuracy of face detection is high, and the algorithm is robust and adaptive. Secondly, face images are affected by uneven illumination. In this paper, a multi-face detection method based on skin color segmentation and feature location is proposed. Firstly, the color deviation is detected in RGB space, and if there is color deviation, the improved reference white algorithm is used to compensate the illumination. Then, the processed images are converted to YCbCr color space for skin segmentation, and the candidate regions are obtained by the improved AdaBoost algorithm for face detection. Then the human eyes are marked in the candidate's face area by the prior knowledge obtained from a large number of human eye training experiments. Finally, the human face image with human eye localization is output.
【学位授予单位】：新疆大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】