基于深度学习及密度峰值聚类的皮肤检测算法研究

发布时间：2018-03-22 20:30

本文选题：皮肤检测　切入点：卷积变换　出处：《南京邮电大学》2017年硕士论文　论文类型：学位论文

【摘要】：皮肤检测技术是众多人体模式识别的基础研究课题之一,它在多个领域中表现出较高的应用价值,例如医学诊断、Web图像过滤、手势识别、面部表情识别等。虽然目前存在众多的皮肤检测算法,但其检测效果仍然存在提高的空间。本文对皮肤检测中的预处理阶段、判决阶段及后期处理阶段进行深入研究,并在此基础上提出改进算法。主要内容及创新点可以概括如下:(1)拍摄过程中光照分布的不均匀性易产生图像像素点之间灰度值的差异,为皮肤检测过程增加了难度。本文提出在皮肤检测预处理阶段引用改进的Z-score标准化与卷积变换相结合的方法来降低光照不均匀的影响,仿真实验结果表明本文算法不仅能够实现对灰度值较低或者灰度值较高的补偿,而且能将整幅图像各像素点的灰度值基本补偿到76-175之间,新的补偿方法提升了皮肤检测效果。(2)传统的皮肤检测方法是通过首先建立皮肤像素点的底层特征信息模型然后验证待检图像像素点是否属于特征信息范围来实现检测的,然而此类检测方法仅仅考虑到皮肤像素点的基础特征信息(如YCbCr颜色空间特征),并没有建立更深层次的皮肤像素点特征,在检测的过程中必然存在一定的局限性。本文提出在皮肤检测技术的判决阶段,使用深度学习算法建立深层次皮肤像素点特征,并使用待检图像与深层次皮肤像素点特征对比的方式来实现皮肤检测,仿真实验结果表明相比其他皮肤检测算法,本文算法在皮肤检测过程中不仅提高了正检率而且降低了误检率,具有较好的性能。(3)在复杂背景图像中,近似皮肤区域像素点会对皮肤检测算法造成较大干扰。本文提出在皮肤检测技术的后期处理阶段通过使用密度峰值聚类分析算法来减弱近似皮肤区域给皮肤检测技术带来的干扰,仿真实验结果表明本文算法在增强皮肤检测技术方面具有积极的影响,不仅减弱了近似皮肤像素点的干扰而且提高了皮肤的检出率。
[Abstract]:Skin detection technology is one of the basic research topics of human pattern recognition. It has high application value in many fields, such as medical diagnosis web image filtering, gesture recognition, etc. Facial expression recognition and so on. Although there are many skin detection algorithms at present, there is still room for improvement in the detection effect. In this paper, the preprocessing stage, the judgment stage and the post-processing stage of skin detection are deeply studied. On this basis, the improved algorithm is put forward. The main contents and innovations can be summarized as follows: (1) the inhomogeneity of light distribution in the shooting process can easily result in the difference of gray values between pixels in the image. In this paper, an improved Z-score standardization and convolution transformation method is proposed to reduce the effect of uneven illumination in the skin detection preprocessing stage. The simulation results show that the algorithm can not only compensate for the lower or higher gray value, but also can compensate the gray value of every pixel in the whole image to 76-175. The new compensation method improves the effectiveness of skin detection. (2) the traditional skin detection method realizes the detection by first establishing the underlying feature information model of skin pixels and then verifying whether the pixel points of the image under inspection belong to the range of feature information. However, this detection method only takes into account the basic feature information of skin pixels (such as YCbCr color space features), and does not establish deeper skin pixel features. There must be some limitations in the process of detection. In the judgment stage of skin detection technology, the depth learning algorithm is used to establish the pixel feature of deep skin. The method of contrast between the image to be detected and the pixel feature of the deep skin is used to realize the skin detection. The simulation results show that compared with other skin detection algorithms, In the process of skin detection, the algorithm not only improves the positive detection rate but also reduces the false detection rate, and has good performance in complex background images. Pixel points in the approximate skin region will cause great interference to the skin detection algorithm. In this paper, we propose to reduce the approximate skin area by using the peak density cluster analysis algorithm in the post-processing phase of the skin detection technology. Interference from technology, The simulation results show that the proposed algorithm has a positive effect on the enhancement of skin detection technology, which not only weakens the interference of similar skin pixels, but also improves the detection rate of skin.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】