Research On Key Techniques Of Content-Based Large-Scale Image Retrieval

With the prevalence of the Internet and digital cameras, there are more and more digital images. In the areas such as commerce, government, academia, and hospitals, digital image is becoming an important approach for information representation. The large-scaledigital images brought two effects, on the one hand, the huge amount of information attracts more and more users, and on the other hand, it is hard for users to find the exact information they really need from the large image database. Therefore, efficient and effective large-scale image retrieval methods have become an important research direction in both academic and commercial circles.The early research mainly focus on the text-based image retrieval technologies and achieve significant improvements with the benefits of existing mature technologies for text retrieval. With the development of research and the increasing scale of digital images, the limitation of early method is more and more obvious. Recent researchers proposed content based image retrieval system, developed image information extraction and indexing methods based on the theories of computer vision, information retrieval and statistical learning. Focusing on the content based large-scale image retrieval problem, we analyze the advantages and disadvantages of the existing approaches, then we take full advantages of computer vision, information coding, and statistical learning to investigate the methods for extracting, compressing and indexing image feature information. The goal is to improve the efficiency and effectiveness of large-scale image retrieval system.The main studies and achievements of the thesis are listed as following:1)Focus on the bag-of-feature based large-scale image retrieval methods, we propose an efficient image retrieval method based on random subwindow feature extraction method and random ferns classifier, and parallel accelerate the method by graphic processing unit. Our method avoid the complex local image descriptorcomputation and high dimensional vector indexing, it is simple, fast and highly paralleled. It significantlyimproves the training and recognition speed when comparing with traditional methods.2)By analyze the advantages and disadvantages of the product quantization based approximate nearest neighbor search method, we propose an efficient approximate nearest neighbor search method based on residual vector quantization, and use it for global descriptor based large-scale image retrieval. It weakens the statistical assumption in the product quantization method and maintains high accuracy both on structured and unstructured vector: the performance on the structured vector slightly outperforms that of the quantization method while the performance on unstructured vector significant outperforms that of the quantization method. We also propose a non-exhaustive search strategy based on the residual vector quantization to improve the search efficiency.3)Focus on the high computation complexity in learning and encoding stages of residual vector quantization method, we design an approximate nearest neighbor search method based on the projected residual vector quantization which combines the advantages of linear dimensional reduction and residual vector quantization. It significant improves the efficiency of learning and encoding with high search accuracy.4)Focus on the disadvantages of transform coding based approximate nearest neighbor search method, we propose a non-exhaustive search method based on transform coding and vector quantization. The effectiveness of our method is demonstrated with several experimental results.