Image Retrieval System Based On Local Features And Visual Context

Recent years, In terms of image information, numbers of images were uploaded by the network users every day which presented an explosively trend. Therefore, it becomes an urgent problem for people to deal with these image data effectively.and establish an image retrieval to help people find their interested images quickly.The current image retrieval technology based on its local features and bag of features and got the good retrieval performance, the retrieval process mainly includes the following areas: local feature extraction, dictionaries building, coding for image features, the computation of similarity. However, due to quantization error of building visual dictionaries and bag of features ignore spatial information strict the precision of retrieval image systems. The paper introduced an image retrieval system based on local features and visual context. And through a series of technology we can improve the retrieval precision and efficiency. Firstly, the author introduced the image retrieval according to local features and visual context; and improved the efficiency through approximate K-Means method for feature space and build inverted index for image codes; Then the author compared the differences in performance due to the different image matching methods; And then analyzed the differences of traditional and improved weak geometrical consistency method in image retrieval; in the last, the paper implement the image retrieval system based the traditional method and the improved method in this paper.The main contributions of the paper included:1. Designed and Finished an image retrieval system based on the local features and visual context information, the system overcome the shortage of considering the local features only in traditional method, the paper combine hamming codes and feature context to re-rank the retrieval and improve the precision.in the addition, the paper use approximate nearest neighbor and inverted index to improve the recall time of the retrieval system.2. Compared the nearest neighbor and approximate K-Means on the searching of the visual dictionary, Through the experiment, the author found that the speed increased more than 100 times under the condition of the retrieval precision less than 1% based on approximate nearest neighbor method. The author also employed the technology of inverse index to improve the retrieval speed by two times in experiment.3. Analyzed several ways of image similarity matching algorithm, including traditional brute force matching, based on visual dictionary and the matching of hamming coding method. The experiments showed that the performance of the brute force matching was good, but required a lot of the retrieval time. The combination of approximate K-Means and hamming coding can solve the problem of accuracy and speed very well.4. Analyzed and compared some visual context methods,including weak geometrical consistency、enchanced WGC、spatial coding GC、strong weak geometrical consistency SGC、WGC with Hamming Encoding and M-WGC、M-EWGC、M-GC、M-HeWgc modified by this paper and so on, Experiments showed these methods modified by us can improve the retrieval precision significantly.