Journal of Computer Applications ›› 2019, Vol. 39 ›› Issue (3): 695-699.DOI: 10.11772/j.issn.1001-9081.2018071588

Previous Articles     Next Articles

Quality evaluation of face image based on convolutional neural network

LI Qiuzhen1, LUAN Chaoyang2, WANG Shuangxi1   

  1. 1. Wuhan Digital Engineering Institute, Wuhan Hubei 430074, China;
    2. School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan Hubei 430074, China
  • Received:2018-08-01 Revised:2018-10-22 Online:2019-03-10 Published:2019-03-11
  • Contact: 李秋珍
  • Supported by:
    This work is partially supported by the "13th Five-Year Plan" Equipment Pre-Research Field Fund (61401320501).

基于卷积神经网络的人脸图像质量评价

李秋珍1, 栾朝阳2, 汪双喜1   

  1. 1. 武汉数字工程研究所, 武汉 430074;
    2. 华中科技大学 计算机科学与技术学院, 武汉 430074
  • 作者简介:李秋珍(1976-),女,湖北浠水人,高级工程师,硕士,主要研究方向:图像质量评价、人工智能;栾朝阳(1991-),男,河南新乡人,硕士,主要研究方向:计算机视觉、深度学习;汪双喜(1989-),男,湖北武汉人,工程师,硕士,主要研究方向:图像处理、深度学习。
  • 基金资助:
    "十三五"装备预研领域基金(61401320501)。

Abstract: Aiming at the low recognition rate caused by low quality of face images in the process of face recognition, a face image quality evaluation model based on convolutional neural network was proposed. Firstly, an 8-layer convolutional neural network model was built to extract deep semantic information of face image quality. Secondly, face images were collected in unconstrained environment, and were filtered by traditional image processing method and manual selecting, then the dataset obtained was used to train the model parameters. Thirdly, by accelerating training on GPU (Graphics Processing Unit), the mapping relationship of fitted face images to categories was obtained. Finally, the input probability of high-quality image category was taken as the image quality score, and the face image quality scoring mechanism was established. Experimental results show that compared with VGG-16 network, the precision rate of the proposed model is reduced by 0.21 percentage points, but the scale of the parameters is reduced by 98%, which greatly improves the efficiency of the model. At the same time, the proposed model has strong discriminant ability in aspects such as face blur, illumination, posture and occlusion. Therefore, the proposed model can be applied to real-time face recognition system to improve the accuracy of the system without affecting the efficiency.

Key words: face recognition, Convolutional Neural Network (CNN), image quality evaluation, quality evaluation of face image

摘要: 针对人脸识别过程中人脸图像质量较低造成的低识别率问题,提出了一种基于卷积神经网络的人脸图像质量评价模型。首先建立一个8层的卷积神经网络模型,提取人脸图像质量的深层语义信息;然后在无约束环境下收集人脸图像,并通过传统的图像处理方法以及人工筛选进行过滤,得到的数据集用以进行模型参数的训练;其次通过在图形处理器(GPU)上加速训练,得到用于拟合人脸图像到类别的映射关系;最后将输入在高质量图像类别的概率作为图像的质量得分,建立人脸图像的质量打分机制。实验结果表明,与VGG-16网络相比,所提模型准确率降低了0.21个百分点,但是参数规模减小了98%,极大地提高了模型运算效率;同时所提模型在人脸模糊、光照、姿态和遮挡方面都具有较强的判别能力。因此,可将该模型应用在实时人脸识别系统中,在不影响系统运行效率的前提下提高系统的准确性。

关键词: 人脸识别, 卷积神经网络, 图像质量评价, 人脸图像质量评价

CLC Number: