人脸识别中眼镜去除方法的研究

发布时间：2018-04-16 14:13

本文选题：眼镜去除 + 眼镜检测　；参考：《河北大学》2017年硕士论文

【摘要】：随着计算机和互联网技术的迅速发展,人脸识别技术受到越来越多研究者的关注。虽然近些年人脸识别技术方面的研究取得了一定的成就,但是仍然受到面部表情、姿态、位置、头部尺寸的变化和遮挡物等因素的限制,其中眼镜作为最常见的遮挡物,会给人脸的后续处理和识别过程带来很大的影响,尤其是结构特点明显的深色眼镜框和与肤色差异较大的“白”色眼镜框。因此,本文主要针对研究意义重大的人脸图像中的眼镜去除进行了研究,本文主要的工作内容如下所示:1.眼镜遮挡区域检测是眼镜去除的前提,本文通过半阈值和最小方差滤波的阈值分割判断是否有眼镜并获得眼镜的遮挡区域。首先用椭圆形及其内部区域代替眼睛来减少其对检测的影响;然后用半阈值分割去掉包含眼睛在内的部分区域,得到包含眼镜在内的显著区域;鉴于眼镜形状的特殊性,为了去除眼镜周围细微的影响从而精确地得到眼镜遮挡区域,采用多个滤波模板的最小方差的阈值分割,最后得到了眼镜遮挡区域。该方法可以较好的去除眼镜外的其他干扰和噪声。2.眼镜遮挡区域相对于人脸图像属于较大的区域,而基于样本块的修复算法可以针对待修复区域较大的图像,所以本文在Criminisi算法的基础上进行改进并用于眼镜去除。首先在优先权中加入与眼睛距离相关的一项,然后根据眼镜遮挡区域检测的中间过程将最佳匹配块的搜索范围减小。该改进方法比单纯的Criminisi算法修复得到的眼镜去除效果更好,减少了眼镜留下的痕迹,提高了人脸识别率。3.由于同一人脸图像中距离较近的皮肤区域差异较小,补偿域之间的距离越小补偿效果越好,因此,本文提出最小距离搜索的扇形线性插值补偿法去除眼镜。首先在眼镜检测的基础上得到初始补偿域;然后为了减少眼镜形状和噪声的影响对补偿域进行适当的调整;接下来根据最小距离原则寻找最佳匹配补偿域;最后根据最佳匹配补偿域使用扇形线性插值补偿的方法得到去除眼镜后的人脸图像,其中对最佳匹配补偿域的快速寻找和扇形区域大小的确定进行了说明。该方法易于实现,眼镜去除视觉效果好,提高了人脸识别率。
[Abstract]:With the rapid development of computer and Internet technology, face recognition technology has attracted more and more attention.Although some achievements have been made in face recognition technology in recent years, it is still limited by facial expression, posture, position, head size change and occlusion, among which glasses are the most common occlusion.It will bring great influence to the subsequent processing and recognition process of human face, especially the dark glasses frame which has obvious structure characteristics and the "white" glasses frame which is quite different from the skin color.Therefore, this paper mainly focuses on the study of the significant face image glasses removal, the main work of this paper is as follows: 1.The detection of glasses occlusion region is the premise of the removal of glasses. In this paper, the threshold segmentation of semi-threshold and minimum variance filter is used to determine whether there are glasses and to obtain the occlusion area of glasses.First, the ellipse and its inner region are used instead of the eye to reduce the impact on the detection; then the semi-threshold is used to segment the area that contains the eye to get the prominent area, which includes the spectacles; because of the special shape of the glasses,In order to remove the subtle influence around the glasses and get the spectacle occlusion area accurately, the minimum variance threshold segmentation of multiple filter templates is used, and finally the spectacle occlusion region is obtained.This method can remove other interference and noise. 2.Compared with the face image, the spectacle occlusion region belongs to the larger region, and the restoration algorithm based on the sample block can be used for the large image to be repaired, so this paper improves on the Criminisi algorithm and applies it to the spectacle removal.First, an item related to eye distance is added to the priority, and then the search range of the best matching block is reduced according to the intermediate process of eyeglass occlusion region detection.The improved method is better than the simple Criminisi algorithm in removing the glasses, reducing the marks left by the glasses, and increasing the face recognition rate. 3.Because the difference of skin region in the same face image is small, the smaller the distance between the compensation regions, the better the compensation effect is. Therefore, a sector linear interpolation compensation method based on minimum distance search is proposed in this paper to remove the glasses.First, the initial compensation region is obtained on the basis of eyeglass detection; then the compensation domain is adjusted appropriately to reduce the influence of glasses shape and noise; then, the optimal matching compensation region is found according to the principle of minimum distance.Finally, according to the method of sector linear interpolation compensation in the optimal matching compensation domain, the face image after removing glasses is obtained. The fast search of the optimal matching compensation domain and the determination of the size of the sector region are explained.This method is easy to realize, and the visual effect of glasses removal is good, and the face recognition rate is improved.
【学位授予单位】：河北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】