A Method on Face Recognition Based on Single Sample Per Person with a Contaminated Biometric Enrolment Database
Jie Li, Chenghu Tang, Deyan Sun · 2020
At present, face recognition stands out among various biometric identification technologies due to its unique advantages, and a lot of face recognition algorithms based on a large number of standard samples have been proposed and applied in real life. In order to deal with more complex scenarios, single sample per people (SSPP) face recognition has become one of the most difficult challenges. In practice, the only remaining sample may be contaminated by illumination, expression, and occlusion, resulting in performance degradation of various existing recognition algorithms, which is called SSPP with a contaminated biometric enrolment database (SSPP-CE). The synergistic generic learning (SGL) method studied in this paper is just a kind of solution based on general auxiliary data set to solve this problem. It uses sparse coding and dictionary learning to extract the less discriminative part (LDP) features of the auxiliary set samples as the variant dictionary, and the more discrimination part (MDP) features of the enrolment database samples as the prototype dictionary. Based on the framework of the prototype plus variant (P+V) model, the method in this paper realizes SSPP-CE face recognition. It is verified that the method has better recognition performance on contaminated samples.However, the existing researches did not identify whether the testing sample belongs to the enrolment database, which will cause misjudgment. To solve the problem, this paper proposes a method based on SoftMax function to identify whether the testing sample belongs to the enrolment database by confidence probability. The experimental results on the AR database prove that when the threshold is set to 0.7 and the contamination degree of the enrolment database is less than 70%, more than 90% of the non-enrolled samples can be identified. This method solves the problem of new testing samples outside the contaminated enrolment database, which improves the overall recognition accuracy for SSPP-CE problem.