The Research On Key Technologies Of Local Sparse Coding Quantization-Based Scalable Face Image Retrieval

In recent years, with the rise of blog and social networking sites as well as the needs of public safety, the face images quickly grow to large-scale, retrieving face images similar to the probe face pictures in such large-scale gallery, has become an urgent demand for users. As a result, the research on large-scale face image retrieval has become a hot spot in research community.For large-scale face image retrieval problem, directly using the traditional face recognition methods, which extract complex and high-dimensional facial features and exhaustive linearly scan the entire face gallery to find the most similar faces, are not scalable. Simplely using state-of-art BOW(Bag-of-Words) model-based algorithms in domain of content-based image retrieval, which ignore the strong specific geometric constraints of the human face, the quantized errors degrade the performance of such system. The state-of-art solutions for large-scale face image retrieval problem, such as Identity-Based Quantization and Multi-Reference Re-ranking, using the priori knowledge of facial local patches’positions information to reduce quantization loss, while using multi-reference set to re-ranking increase the accuracy, achieved good results, however, these solutions1) underutilize face-specific geometric constraints,2) quantization losses and errors are still large,3)the genertatation of reference sets is not robust to age and other changes,which may introduce dissimilar reference faces, therefore, the performance of large-scale face image retrieval can be further improved. For this purpose, we addressed the problem in three aspects, from which further solutions have been proposed.Firstly, at facial feature extraction stage, the combination of local grid patches and global geometry representation, can retrieve the gallery faces which not only local but also general similar to probe face, exploit more special geometric properties of faces.Secondly, for scalability, during the quantization process, using of age and gender constraints extend visual word’s informations, utilizing sparse representation and nearest neighbor search algortithms to quantize local and global facial fetures, reduce the quantized losses and errors which improve the retrieval performance.Finally, the construction of the reference set may introduce dissimilar reference faces by age and other factors, especially in the early iteration. Thus we add semantic classification constraints to the original distance metric, propose a new classification constraints-based multi-reference re-ranking algorithm to improve the performance of the final face image retrieval.The experimental results show, as some face images whose holistic representation are closer but dissimilar have been reject during the process of feature quantization-base retrieval, our proposed approach outperforms the exhaustive linear scan the whole database using global features. Furthermore our method outperforms the state-of-the-art method using Identity-Based Quantization and Multi-Reference Re-ranking on a668,374face database.