Research On Saliency Region-Oriented Image Retrieval Technology Based On Multiple-Feature Fusion

With the development of the Internet and information acquisiton technology, the number of images has grown explosively as the information carrier. In order to find the needed image quickly and accurately, Content-Based Image Retrieval(CBIR) is gradually on the rise. Due to its advantage of versatility, high efficiency and independence from the professional knowledge, CBIR has become the hot issue of the domestic and foreign institutions.This paper focuses on saliency detection, feature extraction and fusion, relevance feedback and has a certain amount of exploration and research. The research work is as follows:(1) Image retrieval systems often treat all the regions of the image equally and result in computational complexity and low retrieval accuracy. This paper proposes a new saliency dectection model which is implemented with a coarse to fine strategy under MRF framework. Experimental results show that the detected section based on this model is more accurate and complete.(2) For one single feature can not accurately describe the image, the paper select Scalable Color Descriptors(SCD), Gabor wavelet and moment invariant as the image feature. Then, based on the selected features, it do the feature fusion using Principal Component Analysis, effectively reducing the feature dimension.(3) There is semantic gap between low-level feature and high-level semantics when we do the retrieval using the low-level features. The paper introduces Echo State Networks(ESNs) to execute the image retrieval. It establishes a good reserve pool and trains the network for image retrieval using the image database. The experiments show that the method well meets the needs of users. (4) To improve the retrieval accuracy and import the user understanding, the paper introduces the Relevance Feedback into the retrieval process. The experimental results show that the integration of user feedback improve the retrival accuracy.