Research On Content - Based Image Retrieval And Recommendation Technology

With the rapidly development of digital image acquisition devices and Internet technology, The number of images in the Internet growth rapidly at a geometric level, this produce big problem for users to find images in the image ocean of Internet. For this reason, using computer extract the image visual content and organization image content, then improve the efficiency of image retrieval and recommendation has became a research hotpot of computer vision research field.This article research the theory and technology related to machine learning based image retrieval and recommendation, including feature extraction, image similarity measure, kernel based image clustering, recommendation image selection and high efficiency visualization and interaction technology and so on, constructed an image retrieval and recommendation system. The main work and innovate of this article is:(1)In order to measure image similarity exactly, this article proposed multi-kernel based image similarity combined measure method. After extract global and local features, construct color histogram, gray level co-occurrence matrix. Gabor and SIFT similarity measure matrix, then combined these matrix into one similarity measure matrix.(2)Because large scale image library has bad linear separability in feature space and clustering performance hardly rely on initial class centre, this article gives an improved kernel based K-means clustering method. After map features into high dimension, then select samples use improved method; use K-means method clustering images in high space. Experiment result shows that this improved method can improve1%accuracy rate compare to original kernel based K-means method.(3)Because large amount of images in image library, it is difficult to display all this images. This article designed hierarchical recommendation method for large scale image library, select representative images for recommendation in all kinds of classes, realize retrieval and recommendation in image library based on user selected recommendation image. Statistical result shows, the method shorten user’s retrieval time.(4)In order to visual large scale image data in two dimensions, this article designed a large scale image visual method based on similarity-preserving projection and hyperbolic coordinate transform. this article adopted kernel based PCA and kernel based MDS obtaining projection point in the two dimension space, use Hyperbolic method transform the coordination properly. Compared to existing method, this method can satisfy user’s demand.(5)Proposed a novel relevance feedback image retrieval method based on features variance, judge user’s like by computer the variance of different features, adjust similarity measure matrix based on variance. Experimental result shows this method improved the accuracy rate.(6)Designed an new image retrieval and recommendation system method, realized recommendation, filter, retrieval of large scale image library. Compared to existing system, our system show particular advantage on retrieval speed and visual method.