Research On Content Based Binary Trademark Image Retrieval

With the development of multimedia technology and fast expanding of visual information, an effective method is sharply needed to manage and retrieve visual information. So, the content based image retrieval(CBIR) technology comes into being and become an important research area in multimedia retrieval and image processing.CBIR is to perform the similarity retrieval according to the image features representing the image content, which may be extracted in the generic or specific domain. With the development of market economy the trademark plays the more and more important role in the current society and the content based image retrieval technique has been widely used in the trademark domain. This paper does extensive and deep research in CBIR technology, and the status quo and key technologies are introduced, the bottle-neck and development trend are also discussed. The paper analyzes the application requirement for CBIR technology in Trademark image retrieval area, presenting three new trademark image retrieval methods and experiments show that the methods keeps good invariance under rotation, translation, scale and the retrieved results match human visual perception very well.Firstly, this paper presents the trademark image retrieval method based on distance distribution information entropy. The image object region is extracted according to the object pixels'circum-circle and then partitioned based on the concentric circles to produce a series of sub-images, according to which the distance distribution information entropy is produced and then normalized. Secondly, this paper presents the trademark image retrieval method based on normalized unit moment of inertia features. Each image is partitioned into numerous unit sub-images, and then the NUMI features of individual unit sub-images of each image are extracted, which are formed into a feature vector for describing the shape of the image. Lastly, this paper presents the trademark image retrieval method based on region orientation information entropy. Image is rotated according to its principal orientation, and the object region on the rotated image is extracted. Then, the object region is partitioned into a lot of sub regions along its circle orientation, and the information entropy of each partitioned regions is computed, which are constructed into a feature vector for describing the shape of the image. Finally, the Euclidean distance is adopted to measure the similarity between the images based on the feature vector of each image obtained. The comparison experiments show that the trademark image retrieval method based on distance distribution information entropy is superior to the one based on distance distribution histogram, and both the one based on normalized unit moment of inertia features and the one based on region orientation information entropy match human visual perception very well.Currently there is no perfect feature for content based image retrieval to make the retrieval result meet all the people's visual perception and all the applications'practical requirements very well. So it is necessary to fuse the multiple features and perform the human-computer interaction in the image retrieval. The three trademark image retrieval methods presented in this paper are almost the same with each other under the evaluation of visual consistency, but the corresponding retrieval results have different styles. All of them are absolutely necessary as the effective alternatives provided to the user in the retrieval system based on multi-feature fusion.