融合生物启发与深度属性学习的人脸美感预测方法

发布时间：2018-04-27 02:37

本文选题：人脸美感分析 + 眼动仪　；参考：《郑州大学》2017年硕士论文

【摘要】：自动人脸美感分析是指通过计算机模拟人类对脸部美感的分析与评价机制,并使计算机不断自主“学习”人类审美方法的一门技术。近年来,随着人脸美感评价需求在美容、婚介、招聘、多媒体等不同行业的不断增加,自动人脸美感分析技术得到了迅速发展。传统的自动人脸美感分析方法致力于找到能够影响人脸美感程度的特征和高精度的分类算法,以完成对不同人脸的美感分类与评价。但这些方法在分析图像信息有效表达的过程中,不能清晰的解释人类视觉系统和大脑是如何筛选和运用这些特征来进行人脸美感判断,缺乏科学的方法去统计和验证其合理性。为了尽可能地还原人类的审美判断过程,得到更好的分类效果,本文提出了融合生物启发和深度属性学习的人脸美感预测方法,一方面使用基于中层特征表示的方法提高图像特征提取算法的性能,另一方面使用深度学习和模型融合的方法提高分类算法的精度,最终构建出基于深度学习和模型融合的人脸美感分析框架,本文的主要工作分为如下几个方面:(一)提出了一种基于图像中层语义的特征提取方法。本方法首先通过眼动仪实验提取人在审美过程中的人脸显著性区域和人脸显著性权重矩阵,然后通过“整体+局部”的视觉注意机制获取决定人脸美感的仿生全局区域,并验证了仿生全局区域的有效性,最后从仿生全局区域中提取图像的中层语义特征。该方法的目的是得到图像中层语义的特征表示,相较于只包含图像基本信息的图像底层特征表示的方法,本方法因其对图像语义的更好理解使其具有更好的表达能力,在分类能力上也具有更好的效果。实验结果表明这种基于图像中层语义的特征提取方法在人脸美感表达中表现出更好的效果。(二)提出了一种基于TrueSkill算法的人脸美感标签确定方法。直接通过被试对样本美感进行打分不够精确,得到样本之间美感的相对排序却比较简单,本文首先建立由相同规格人脸图像组成的样本集,通过美感排序实验得到被试对每组样本图像的美感排序结果,然后通过TrueSkill算法将相对美感排序转换为绝对的美感分数,最后通过加权平均得到人脸美感标签。(三)提出了一种基于深度学习和模型融合的人脸美感分析框架。本框架首先通过卷积神经网络训练出一系列基于仿生全局特征的仿生属性检测器,并得到训练样本的概率输出,然后将每个训练样本在不同检测器中的概率输出融合为该样本的特征向量,最后结合美感标签训练出人脸美感预测模型。本框架得到的人脸美感预测模型提高了分类算法的准确度。实验结果表明,本框架基于深度学习和模型融合的方法不仅相比于单个仿生属性检测器具有更高的分类准确度,相比于其它的特征提取方法也表现出了不俗的效果。
[Abstract]:Automatic face aesthetic analysis is a technique that simulates the analysis and evaluation mechanism of human face aesthetic perception by computer, and enables the computer to "learn" the human aesthetic method. In recent years, with the increasing demand of face aesthetic evaluation in different industries, such as beauty, matchmaking, recruitment, multimedia and so on, automatic face aesthetic analysis technology has been developed rapidly. The traditional automatic face aesthetic analysis method is devoted to finding the features and high precision classification algorithms that can affect the aesthetic degree of the face in order to complete the classification and evaluation of different faces. However, these methods can not explain clearly how the human visual system and brain screen and use these features to judge the aesthetic perception of human face in the process of analyzing the effective expression of image information, and lack of scientific methods to statistics and verify its rationality. In order to restore the process of human aesthetic judgment as much as possible and get better classification effect, this paper proposes a face aesthetic prediction method which combines biological heuristics and depth attribute learning. On the one hand, the method based on middle level feature representation is used to improve the performance of image feature extraction algorithm; on the other hand, the method of depth learning and model fusion is used to improve the accuracy of classification algorithm. Finally, a face aesthetic analysis framework based on depth learning and model fusion is constructed. The main work of this paper is as follows: 1. A feature extraction method based on image middle level semantics is proposed. In this method, the salience region of human face and the weight matrix of face salience are extracted by eye movement experiment, and then the bionic global region which determines the aesthetic sense of face is obtained by the "global local" visual attention mechanism. The validity of the bionic global region is verified. Finally, the middle semantic features of the image are extracted from the bionic global region. The purpose of this method is to get the feature representation of image middle level semantics. Compared with the method of image bottom feature representation which contains only the basic information of image, this method has better expression ability because of its better understanding of image semantics. In the classification ability also has the better effect. Experimental results show that the feature extraction method based on image meso-semantic is more effective in facial aesthetic expression. (2) A face aesthetic label determination method based on TrueSkill algorithm is proposed. It is not accurate enough to score the aesthetic feeling of the sample directly, but it is relatively simple to get the relative ranking of the aesthetic feeling between the samples. In this paper, a sample set composed of face images of the same size is first established. The aesthetic ranking results of each sample image were obtained by aesthetic sorting experiment, then the relative aesthetic sorting was transformed into absolute aesthetic score by TrueSkill algorithm, and the face aesthetic label was obtained by weighted average. (3) A face aesthetic analysis framework based on deep learning and model fusion is proposed. Firstly, a series of bionic attribute detectors based on bionic global features are trained by convolution neural network, and the probabilistic output of training samples is obtained. Then the probabilistic output of each training sample in different detectors is fused into the feature vector of the sample. Finally, a face aesthetic prediction model is trained by combining the aesthetic label. The prediction model of facial aesthetic perception obtained by this framework improves the accuracy of the classification algorithm. The experimental results show that the method based on depth learning and model fusion not only has higher classification accuracy than the single bionic attribute detector, but also has a good effect compared with other feature extraction methods.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】