基于特征偏移的人脸图像质量评价及影响因素研究

Face image quality evaluation based on feature offset and research on influencing factors

摘要:

针对低质量人脸图像阻碍识别系统性能提升的问题，本文提出了一种无参考的人脸图像质量评价方法，并使用该方法评估了不同类型的图像退化对人脸图像质量的影响程度。该方法使用一种集群卷积网络结构，模拟人脸图像退化过程中的特征偏移，根据特征偏移量和图像信息量之间的相关性，完成人脸图像质量分数计算。使用遗传算法对构成集群网络的网络单元进行筛选，可使用更小网络规模实现同等性能。以质量评价算法为工具进行实验，研究评估了不同图像退化类型对人脸识别的影响，为指导今后人脸质量相关研究得出了有益结论。在主流人脸数据集上进行的实验证明，通过筛选数据库中低质量分数的人脸图像，可以进一步提升现有人脸识别系统的性能，且识别率提升表现出良好稳定性。该方法复杂度低，无需训练，与FaceQNet等最新方法相比，在FNMR和EER指标上显示出明显优势。

Abstract:

Aiming at the problem that low-quality face images hinder the performance improvement of recognition systems, we propose a no reference face image quality evaluation method to evaluate the impact of different types of image degradation on the quality of face images. This method uses a cluster convolutional network structure to simulate the feature offset in the face degradation process, and face image score is calculated based on the correlation between the feature offset and the amount of image information. The genetic algorithm is used to select the network units, so that the same performance can be achieved with a smaller network scale. Using the evaluation algorithm as a tool for experiments, the impact of different image degradation types on face recognition is evaluated and studied, some useful conclusions are drawn for guiding future research on face quality. Experiments conducted on mainstream face data sets show that by filtering low-quality face images in the database, the performance of the existing face recognition system can be further improved, and the improvement of the recognition rate shows good stability. This method is low in complexity and does not require training. Compared with the latest methods such as FaceQNet, it shows obvious advantages in FNMR and EER indicators.