Research And Implementation Of Face Retrieval Method Based On Internet Cafe Surveillance

Face recognition is a feature that is quickly gaining attention in the industry, mainly due to its naturalness and oblivious. This feature is already available for practical applications under controlled conditions. There are many face recognition products on the market and the number is continuously growing. With the development of Internet and the construction of the safe city, face recognition is being used widely and in different sectors. Demand is continuously increasing and changing and it has become difficult for the producers to be able to meet the market requirements. There is especial a need to come up with an algorithm that works, under natural conditions, such as cases where one is affected by illumination and posture.This dissertation is aimed at structuring a Face Retrieval System, based on cyber cafe video surveillance, which is used to face pictures retrieval in video and collect identity in real-time. We used the tools in different scenarios and explored face feature extraction and retrieval algorithms with excellent performance on the basis of existing face collection and retrieval methods.The main contents and achievements are as follows:1. Considering sparse samples, this dissertation proposed a facial features extraction algorithm with multi-features fusion. At the beginning of deployment, the system can’t obtain enough face images for training, because of which, the effect of current frequently-used feature extraction method based on deep learning decreases significantly. Although we can get great performance with traditional artificial features in certain cases, they are, however, unqualified generally. Taking these into consideration, to find a facial features extraction algorithm with multi-features fusion is becoming the way to solve this problem. Experiments show that compared to single feature retrieval method, accuracy rate of the above method increased significantly.2. Considering dense samples, this dissertation proposed a facial retrieval method with relevance feedback which is based on Deep Neural Network. As system deployment time goes, number of samples for training becomes larger and larger. With a huge number of samples, it becomes important to further improve the face retrieval system’s performance. This dissertation finished the extraction of facial features with Convolutional Neural Network trained by a large number of training samples, and trained the feedback model with the data collected from others (feedback) about retrieval results. This makes the retrieval accuracy rate higher than other introduced methods and makes system more intelligent during the accumulation of feedback samples.3. Combined with existing face detection algorithms, we applied the above facial feature extraction and retrieval algorithm to the face detection and retrieval system which is named "Wise Eyes". The system has been deployed in Hefei Public Security Bureau as a product with a good market prospect.