Research On Region-based Image Retrieval System With Fuzzy Features

Region-based image retrieval(RBIR) is an important research branch of CBIR. In RBIR, an image is segmented into regions by image segmentation, whose visual features are used to represent and index images, which can reduce the gap between low-feature and high-level semantic concepts and be closer to the human perception.Image retrieval is related to fuzzy set theory naturally. The paper brings fuzzy set theory into RBIR, and does research into image segmentation, fuzzy features extraction for regions, similarity computation, index tree building and relevant feedback and so on.As for region-based image segmentation, we segment image into regions with K-means cluster algorithm for 4*4 pixel blocks, and propose a new method that is based on pixel block features to choose the best cluster value of K with derivative. Experiment results indicate that our method can segment images efficiently. As for region fuzzy features, we apply fuzzy set theory to obtain fuzzy features for color, texture, shape with Cauchy membership function, which enhances image segmentation robustness and image retrieval effects.As for image similarity computation, we present an integrated region match approach which firstly computes respectively region integrated features(color, texture, location), histogram, shape, then treats each similarity product with weight as similarity between two regions, finally similarity between two images is the means of each region maximum. Our method improves the retrieval precision, and has better performance, which is validated by some experiments.As for index tree building, we propose an improve cluster method to build index tree that is based on K-means cluster algorithm with initial clustering for cluster number and cluster centers, and search algorithm with A* search algorithm, N-neighborhood and triangle inequality, which improves the retrieval speed.As for relevance feedback, we introduce an SVM-based image query method, which make the system guess user's interest by his choices and automatically adjust the measurement rule to improve the accuracy. SVM theory classifies features vectors sets, and builds relevant feedback mechanism by pruning index tree to session tree, which makes retrieval effects closer to users.Based on the above researches, we design and implement a region-based image retrieval with fuzzy features system, called RBIRFFS. In our proposed system, the feature database is constituted with SQL Server to save image features, and some function modules are included, such as feature extraction, image segmentation, image retrieval, index tree building, relevance feedback and etc, and it has certain practical values.