Research And Application Of Large Scale Face Image Retrieval

With the development of Internet technology, nowadays we are in an era of information explosion. How to effectively organize and utilize these information, has become one of great challenge in modern times. The image of its own resources, contain rich, useful information, the information cannot be directly processing machine. Therefore, a large-scale image retrieval has become a great research significance and theoretical value of the subject. Face image retrieval as a specific large-scale application of them, is also a research direction of the current popular.Face image retrieval, can be divided into two branches, namely the design of face image coding and high dimensional indexing. A good facial image coding not only to distinguish the difference between different people, and different posture, facial expression of the same person has certain tolerance. Problem of high dimensional index design faces is "dimension disaster" problem and the retrieval efficiency problem. Increase with increase of vector dimension and the scale of the data, the performance of face image retrieval system will drop sharply. In this paper, the face image retrieval in large-scale as the research subject, aims to achieve large-scale one million level face image retrieval system. The main contents are as follows:1. Facial image coding. A face image with ordinary objects image is different, it is a kind of non-rigid, has a special nature. Facial features of the face is obvious, and the relative fixed position. Based on these analysis, this paper selects five key reference points in face images(left eye, right eye, nose, mouth, left right mouth), and to extract the local features of the corresponding. On the whole face image, we use Denoise Autoencoder automatically generate global feature.2. Design of high dimensional index. The French Institute of automation, Jegon in 2010 proposed asymmetric distance calculation of inverted index mechanism(Inverted File with Asymmetric Distance Computation, IVFADC), the product quantization asymmetric distance calculation inverted index mechanism is very good(Product Quantization, PQ), non-symmetric distance calculation(Asymmetric Distance Computation, ADC) and inverted index(inverted index) organic fusion. The realization of high dimension, efficient retrieval of a large data set. In this paper, on the basis of IVFADC, different quantization of the global features and local features, and in the establishment of inverted index in the process, for each input vector matching two nearest neighbor clustering center, which increase the half space consumption, in exchange for the whole system of recall.