Research On Content-Based Image Retrieval Technology

With the development of the computer technique and the appearance of large numbers of digital products, there are more and more digital images. How to retrieval image efficiently and accurately that become an important topic. Content based image retrieval includes extracting the feature of images automatically and calculating the similarity distance of the features to retrieval images that we expect.The paper introduced the development of the content based image retrieval at home and abroad, and elaborated the main hot spots and difficulties of this technology at present. Moreover, this paper mainly researched theory basis knowledge and related technology of image retrieval. Based on the current problems for image low-level features and similarity measure, the paper mainly studied several methods as follows:Firstly, if images were retrieved using only a kind of features, there were problems with limitations. So this paper proposed an image retrieval method based on color histogram and contour extraction. Color histogram was extracted in HSV color space. The quantization error of H,S,V color components was reduced using the corresponding mean value of R,G,B that were not quantized. The similarity measures were based on Gaussian blur function. The image contour was extracted based on the combining of improved Canny edge detection and region segmentation. If image contour was clear and accurate, hu-invariant moments feature was extracted better. The experimental showed that this method had better retrieval performance than method in this literature, especially for images with highlight objectives.Secondly, in order to resolve the difficulties in image segmentation, this paper proposed an image retrieval method based on connected region color and texture features. The division of the images simply can embody the shape features of images. And color and texture features were extracted in the connected regions. Texture features were extracted based on gray-level co-occurrence matrix, color features were extracted based on color histogram that can embody color information and hue co-occurrence matrix can embody space distribution information. And dynamic region matching was used on similarity mesure. It could overcome the problem of not rightly matching between these regions. The experimental showed that this method can have the better retrieval performance than the method using one or two bottom visual features of images. Finally, for some methods proposed in the paper and design requirement of retrieval system currently, this paper built an image retrieval system. The system was simple and can be operated by many users. Moreover, the system had higher precision and recall. Especially, the system had expansibility, and importantly introducing the system can be used on trademark and medical images field.