Study And Implement On Large-Scalable Face Image Fast Retrieval

With the exponential growing of large-scalable face image on the internet, a fast and efficient face image retrieval method has shown to be more and more important. How to effectively process and retrieve large-scalable face image has become a current research focus.The main problems of large-scalable face image retrieval include:(1) First, fast large-scalable high-dimensional vector retrieval problem;(2) Similarity, fast large-scalable face image retrieval problem. The main objective of this paper is to address these issues. Specifically, the main work and achievements are summarized as follows:1. Refer to the large-scalable high-dimensional vector retrieval problem we propose two methods. First, we propose a retrieval method based on distributed LSH. In this proposed method, we treat the feature of video files as input of high-dimensional vector and map eigenvalues by applying hash function. With those mapping eigenvalues, we can build an index. The similarity of the relevant videos is calculated according to the returned similar frame sets. Further, we propose another approach based on SimHash. In this method, we compute the DCT spatiotemporal feature of key frames. Given a video, its signature can be generated by applying Simhash to all its key frames. Similar video query is achieved by using efficient hamming distance retrieval. Experimental results show that the average accuracy and average recall rate of retrieval for a similar video in a video database of ten thousands videos is great than90%and the average query time is less than0.11second.2. In order to retrieve the desired face image from large-scale face image database, we propose an efficient fast method for similar face image retrieval. Firstly, we extract LBP features of face images and do dimensionality reduction by mapping the features from Euclidean space into Hamming space. Then a signature for each image is constructed by encoding dimensionality reduced features, using enhanced multi-bit encoding method. The similarity between each signature is judged by Manhattan distance instead of Hamming distance. Finally, we construct inverted indexes from image signatures and fast retrieval is accomplished by using efficient inverted indexes. The experiment on dataset containing200,000face images shows that the average precision rate is78%and the average retrieval time is less than0.15seconds.3. We design an internet bar staff monitoring system which consists of client and face image retrieval module as the face retrieval application. In the design of face retrieval model, we use the face standard dictionary as the visual words quantization standard of the face image database. We need firstly, partial the face and extract the feature of each part. Then we do identity-based quantization, which forms the visual words, to each region using face standard dictionary based on nearest neighbor condition. Hence, we can represent face by visual words. Further, we build the inverted index for visual words to achieve a rapid retrieval for similar face images. Eventually, the ultimate realization of internet cafe staff monitoring system is simple and convenient, user-friendly, search results accurate. The system is under beta testing in internet bar of Hefei now.