Large-scale Face Image Coding And Application In Face Verification

With the rapid development of multimedia technology and explosive growth of image information, the scale of image database is becoming increasingly vast. However, images which contain faces have drawn doubled public attention. In the arena of public security, information retrieval and social networking, it is becoming increasingly significant to address how to manage these massive amounts of face images effectively and efficiently. Large-scale face image coding is crucial to realizing face image retrieval and face verification and is turning into the major breakthrough in enhancing the performance of face retrieval and verification.This paper aims to realize large-scale face image coding under the circumstance of large-scale face image retrieval and face verification. We have proposed some methods to improve the performance of large-scale face image representation through three aspects, including face image pre-processing, face image feature extraction and face image quantization. This paper includes main studies on:(1) We propose an improved explicit shape regression algorithm to realize face alignment in the face image pre-processing stage. This algorithm improves the setting of the initial image and the selection of correlated features. The experiment indicates that the improved face alignment has better results in face feature points location, which offers better support for face region partition and face image feature extraction.(2) Then, in the stage of face image feature extraction, we mention relative face attribute features. We use sequential learning method to quantify face attributes on the basis of face binary attributes and prove through experiment that relative face attribute features could describe face attributes better than face binary attributes.(3) Finally, in the stage of face image quantization, we conduct sparse coding through using the combination of face image’s local features and relative face attribute features. Therefore, we could realize the underlying characteristic of similarity requirements and avoid the problem of ’Semantic Gap’ through introducing face attribute features simultaneously.We have applied the coding method in this paper to the experiment of large-scale face retrieval and face verification. This shows that the coding scheme mentioned in this paper could get better retrieval results and verification accuracy in the comparison of underlying characteristics of coding and attribute feature coding.