Research On Image Retrieval Based On Feature Fusion And Inverted Multi-Index

Image retrieval, an important research field on computer vision, has a broad market demand in a lot of application areas, such as medical diagnosis, organization and administration of digital media, registration management of patents and trademarks, etc. Conventional image retrieval methods based on text labeling have some disadvantages, such as heavy overhead on manually labeling, matching deviation due to individual subjectivity, are incompetent to accomplish large scale image retrieval tasks, while content based image retrieval (CBIR) is playing a more and more significant role in real world. However, content based image retrieval is also faced with some challenges and problems so far, such as semantic gap problem, high dimensional indexing mechanism problem, etc.This thesis improves conventional Color Names (CN) feature, proposing a new color feature called cumulative Color Names (cCN), to alleviate the sparsity problem inside the conventional CN color feature vectors. In consideration of the cohesive relation between vector components, the presented cCN color feature accumulates the probability distribution values for different color components, eliminating the sparsity issue in the conventional CN color feature vectors, reducing the influence of noise and illumination variance, and increasing the stability of visual matching. In addition, the proposed cCN feature makes results based on geometrical distance measurement in accord with similarity results from human visual perception. Besides, this thesis incorporate RootSIFT, cCN color feature and Gabor texture feature for image retrieval, according to the shortcomings exposed in image retrieval based on single image feature, such as partial reflection of image content characteristics, limited matching power, low retrieval precision, etc. What’s more, we proposed a dynamically updating adaptive weights allocation algorithm (AWAA), in order to take full advantage of image features, allocating reasonably different fusion weights. During feature fusion, AWAA adjusts fusion weights automatically, and works out the optimal weights for feature fusion, based on different representation ability of image content, and image content characteristics in different image datasets. This algorithm can improve the interaction between human and system, increasing the retrieval accuracy.Finally, this thesis proposes a new retrieval framework based on multiple features fusion and inverted multi-index (mFFMI), to deal with the problems exposed in conventional inverted index. For a start, constructing a three-dimensional inverted multi-index retrieval framework, in which each dimension of mFFMI is one kind of image feature. Train codebooks of different sizes for each image feature, respectively, and we conduct feature fusion at indexing level. Under Bag-of-Words (BoW) model, the search space is subdivided more densely by codebooks trained in advance. With the combination of different image features, the accuracy and robustness of image matching can be improved.Extensive experiments on several benchmark datasets, such as Corel-1000, Ukbench and Holidays, etc., demonstrate that, the proposed methods in this thesis improve the matching power of image features, with favorable retrieval results in accord with the visual perception of human beings. These approaches bring significant boosts on both precision and recall, enhancing the retrieval performance.