Research And Implementation Of Image Retrieval Based On Multi-Features Fusion

With the development of mobile internet and social network, picture and other multimedia information have become mainstream, the majority of users are used to share their latest development by photos, which makes the number of images on the internet for explosive growth. The development of information technology makes users around the world can access information in different ways, which further promotes people a huge demand for digital images. In face of such a mass of images, traditional image search method by text annotation could not meet people’s demand anymore. How to store image efficiently and retrieve image quickly has become an hot topic in the field of multimedia information retrieval.This thesis mainly focused on the image feature extraction and indexing procedure of the content based image search. At first systematically discussed the image low-level visual feature extraction technology, and then deeply analyzed the image indexing and image matching method. In this thesis, the main research jobs and contributions are as follows:Firstly, doing single feature image retrieval experimental comparison between the common color, texture global features and SIFT, SURF local features on Holidays dataset, by analyzing of their feature extraction time, feature indexing time, feature retrieval time and their mean average precision. In order to make full use of global feature and local feature’s advantage, I designed a feature fusion extraction algorithm combined with the local SURF feature and the global CEDD feature,. The experimental result showed that this method can keep the image extraction efficiency. Besides the feature fusion image retrieval method’s mAP result is better than single feature retrieval way. The feature fusion retrieval method can effectively improve the image feature’s discrimination and robustness. Afterwards by using the classical Bo VW model to build the inverted index structure and save to the lucene index file format, which can further speed up the search time. Besides by adding Hamming Embedding, Soft-Binning procedure, which can further improve the search result and make it able to cope with the requirement of the large dataset index. Finally, this thesis designed and implemented a feature fusion JavaWeb retrieval system, named "One Click Search", through experimental test verified the feasibility of my method. Last but not least, the thesis summarized the main work and its insufficiency, and proposed possible directions of further study.