An Effective De-noising Algorithm for Making A Large Celebrity Face Dataset with A High Purity

摘要

In recent years, face recognition has been greatly improved by the development of CNN such as DeepID, FaceNet, and so on. However, the performance of those trained models is not satisfactory when applied on Asian face recognition because those models are almost trained on western face datasets. In order to solve such a problem, we create a large Chinese celebrity face dataset, including 11,289 celebrities and 395,130 face images, and then we can make a fine tune on those trained models for Asian face recognition by our created Chinese celebrity dataset. In this paper, we make a scheme to collect celebrity names and images from unstructured data by internet. Then we propose an effective de-noising algorithm to improve the quality of dataset, and the purity of our data can reach 97.7% from original 65.9% after the de-noising. Meanwhile, the de-noising operation on MS-Celeb-1M has been realized for evaluating the proposed method, and the purity of the tested fine part of MS-Celeb-1M has been improved from 73.0% to 98.7%. Therefore, the experiments on our created Chinese celebrity face dataset and MS-Celeb-1M indicate that the proposed de-noising algorithm has achieved excellent performance for improving the quality of dataset.