摘要
对于在正常照明技术下采集得到的人脸图像,现有的人脸识别算法如MTCNN算法、RetinaFace算法,已经能够取得相当高的人脸识别率。然而在某些特殊应用中,对于在特殊照明技术下采集得到的人脸图像,现有的人脸识别算法是否具有很好的鲁棒性,保持较高的人脸识别率,并没有确切的实验结果能够给出我们结论。本文收集了6 000多张在不同照明技术下得到的人脸图像作为我们的混合人脸数据库,并利用LBPH算法、卷积神经网络(CNN)、MTCNN算法、RetinaFace算法设计出了四种有效的人脸识别网络,分别对开源的WIDER FACE人脸数据库和我们的混合人脸数据库进行了测试。最终发现RetinaFace算法对于不同照明技术得到的人脸图像具有较好的鲁棒性。我们进一步利用深度学习标注工具对RetinaFace算法误判的人脸图像进行了标注,并将标注后的图像送入到RetinaFace人脸识别网络中重新训练,优化后的RetinaFace人脸检测模型得到了98.6%的人脸识别准确率,使得RetinaFace算法对不同光照条件的鲁棒性取得了进一步的提升。
The existing face recognition algorithms, such as MTCNN algorithm and RetinaFace algorithm, have been able to achieve quite high recognition rate for the face images collected under normal lighting technology. However, in some special applications, there is no exact experimental result that can give us a conclusion on whether the existing face recognition algorithms have good robustness and maintain a high recognition rate for the face images collected under special lighting technology. We collected more than 6 000 face images obtained under different lighting technologies as our hybrid face database, and choose four effective face recognition networks designed by using LBPH algorithm, convolutional neural network(CNN),MTCNN algorithm and RetinaFace algorithm, which are tested on the open source WIDER FACE database and our hybrid face database respectively. Finally, it is found that RetinaFace algorithm is robust to face images obtained through different lighting technologies. We further use the deep learning labeling tool to label the face images misjudged by RetinaFace algorithm, and send the labeled images to RetinaFace face recognition network for retraining. The optimized RetinaFace face detection model achieves 98.6% recognition accuracy, which further improves the robustness of RetinaFace algorithm to different lighting conditions.
作者
孙金龙
吴振宁
肖仲喆
黄敏
SUN Jinlong;WU Zhenning;XIAO Zhongzhe;HUANG Min(School of Optoelectronic Science and Engineering,Soochow University,Suzhou Jiangsu 215006,China)
出处
《电子器件》
CAS
北大核心
2022年第5期1123-1128,共6页
Chinese Journal of Electron Devices
基金
国家自然科学基金项目(61906128)
江苏省自然科学基金项目(BK20180834)。
