Research On Weakly Random Mapping Based Object Retrieval Technologies

In recent years, with the fast development of Internet and multimedia techniques,multimedia data grows explosively and a massive information environment is generated. Facingthe massive image data, people are always interested in some specific objects, but not the wholeimages. Therefore, how to retrieve the specific object accurately and efficiently from alarge-scale image dataset is becoming a very important issue. The paper mainly researches onobject retrieval techniques, the contributions are listed in the following three aspects:(1) In object retrieval area, the current mainstream solution is BoVW (Bag of Visual Words)method. However, there are several problems existing in the conventional BoVW methods, suchas low time efficiency, large memory consumption and the synonymy and ambiguity of visualwords. Therefore, a weakly randomized visual dictionaries model is proposed in this paper.Firstly, E~2LSH (Exact Euclidean Locality Sensitive Hashing) is used to cluster SIFT features ofthe training dataset. Then, the selecting process of hash functions is effectively supervisedinspired by the Random Forest ideas to reduce the randomcity of E~2LSH. Finally, a group ofweakly randomized visual dictionaries is generated with the multi-hash tables to further enhancethe distinguishability of dictionaries. Experimental results show that the cluster result and therepresentativeness of visual words is effectively improved compared with the classical clustermethods, besides, it can increase the efficiency.(2) The weakly randomized visual dictionaries model is combined with query expansiontechnique and an object retrieval method based on weakly randomized visual dictionaries andquery expansion is proposed. Firstly, the SIFT descriptors of image dataset and query object aremapped to the nearest neighbor visual words using E~2LSH to obtain frequency vectors, which arethen weighted according to tf-idf strategy to construct the visual vocabulary histograms. Andthen, the visual vocabulary histograms are stored on disk as index files to overcome the largememory consumption problem. Finally, an average query expansion strategy is introduced toenrich the information captured in original query object region and further improve retrievalperformance. Experimental results indicate that the object retrieval accuracy of the novel methodis boosted dramatically and we can obtain more images which contain the query object comparedwith the classical methods; besides, it adapts large scale datasets well.(3) To effectively utilize the spatial information among visual words and the contextualinformation of object region, the weakly randomized visual dictionaries model is then combinedwith contextual semantic information. An object retrieval method based on weakly randomizedvisual dictionaries and contextual semantic information is proposed. The space position of visualwords is added to the proposed method on the basis of language model, and then, new objectmodel consisting contextual semantic information is devised, which is drawn from the visualelements surrounding the query object. In the end, Kullback-Leibler divergence is introduced asa similarity measurement to accomplish object retrieval. Experimental results demonstrate thatthe new method is useful for disambiguating visual objects which are cluttered, blurred, or occluded and can further enhance the object retrieval performance.In conclusion, in aspect of modeling, the discriminability and distinguishability of visualdictionaries generated is enhanced by expanding and improving randomized mapping. While inaspect of key techniques, we introduce query expansion and contextual semantic information andcombine with weakly randomized visual dictionaries model which can form new object retrievaltechnology and improve the robustness of models and relative algorithms. Moreover, it can makethe object retrieval system still keeping a well performance in complex environments.