基于CNN的人脸图像亮度和清晰度质量评价

为提高人脸识别的准确率,提出基于CNN的人脸图像亮度和清晰度的质量评价方法,拒绝低质量人脸图片参与识别和认证。利用人脸识别匹配器的相似性分数与人类视觉系统清晰度等级分类方法,结合传统亮度分级方法,将人脸图像分成9类并建立相应的数据标签,基于数据标签和数据集训练CNN模型并进行实验。实验结果表明,模型在人脸图像质量的亮度和清晰度方面有了较为准确的评价,准确率达到90%以上。EVR曲线整体FNMR值变化明显,验证了分类的有效性,评价结果对人脸识别系统回路的动态调整有实际指导意义。

Quality assessments of illumination and sharpness of face image based on CNN

To improve the accuracy of face recognition,a quality evaluation method of face image brightness and sharpness based on CNN was proposed to reject low quality face images to participate in recognition and authentication.The similarity score of face recognition matcher and the definition grade classification method of human visual system were used.Combining with the traditional brightness classification method,the face image was divided into nine categories and the corresponding data labels were established.The CNN model was trained and experimented based on data labels and dataset.Experimental results show that the model has more accurate evaluation of the brightness and sharpness of face image quality,and the accuracy is more than 90%.The overall FNMR value of EVR curve changes obviously,which verifies the effectiveness of classification.The evaluation results have practical guiding significance for the dynamic adjustment of face recognition system loop.