Technology Research, Content-based Image Retrieval

Content-based image retrieval(CBIR)is a technique for retrieving images on the basis of image features such as color,texture and shape,etc.Key issues in CBIR include extracting features from raw images,matching query and stored images in a way that reflects human similarity judgement. The paper researches how to derive features automatically and how to match image perceptual similarity as well as possible.First,the paper presents a novel algorithm to extract color signatures. The algorithm exploits the fact that human vision perception is more sensitive to the changes in smooth regions than in detailed regions. To use it,a exponential of color complexity is assigned to every image,Individual color signature of every image is calculated based on image content.Second,a full automatic classification of raw images into textured and non-textured images is presented for large-scale image databases. The algorithm uses region segmentation and statistical testing. As a result,textured and non-textured image will be processed in different ways on image retrieval.For textured image,a color signature is extracted using the uniform distribution. At the same time,an energy-distribution signature is calculated using Gabor filters. Content-based textured image retrieval system establishes color-based and shape-based retrieval. In different queries,the results will be got according to similarity measurement based on different signature.For non-textured image,image retrieval is divided into whole image retrieval and part image retrieval. In part image retrieval,the query results are returned by estimating the scale at which a query image appears within an image. In whole image retrieval,established features include a color signature,the mean values and center points of color R. G. B.,an energy-distribution signature by Fourier polar transform. The results are returned in hierarchical retrieval steps.In similarity measurements mentioned above,the paper analyses the drawback of bin-by-bin-form distance and cross-bin-form distance. The best-weight distance is presented which matches human vision as well as possible. Also,the principle of the distance is based on transport problem. The experimental tests the distance is efficient in image retrieval.In the end,the further research direction is pointed out.