Text And Content-based Trademark Image Retrieval

Content-based image retrieval is a comprehensive study include mathematics, electronics, computer science and artificial intelligence. It is mainly about how to carry on the similar retrieval as the index based on the image content characteristics such as color, texture, shape and spatial distribution. As an important symbol of products, the trademark represents the quality of products and the manufacturer prestige and plays the vital role in the market. And the retrieval technology based on the image content is widely used in the field of trademark image. So the trademark image retrieval actually is a subset of image retrieval. This paper firstly analysizes the method of trademark image retrieval based on the content; and then it proposes a new method of trademark image retrieval; finally it designs and develops a trademark image retrieval system based on the characteristic fusion .For the trademark images, the retrieval is usually developed by the shape characteristic. We can use the moment characteristic to express the shape statisticals. In the image processing, the moment may be regarded as an important characteristic to describe the object, which is often used in some operations of image retrieval. This paper contrast some retrieval methods such as the shape characteristic retrieval (Hu invariant moments, Legendre moment, the Zernike moment, the false Zernike moment) and fuzzy direction characteristic retrieval. and the PVR value is used as the image retrieval performance evaluation criterion. The result indicates that the Legendre moment, the Zernike moment and the false Zernike moment retrieval are better than the Hu moment in aspect of shape characteristic retrieval; the retrieval based on the fuzzy direction characteristic is better than the Hu moment. the false Zernike moment retrieval works best under the same dimension.Along with moment exponent number and characteristic dimension increase, some problems of information redundancy and dimension disaster will be brought in although the image description ability can be improved. In addition ,the higher order moment is more easily influnced by noise than the low-order moment. So the retrieval effect is not proportional to the moment exponent number. Through the experiment, we have discussed the best dimension of the Legendre moment, the Zernike moment, the false Zernike moment and the fuzzy direction characteristic.Because the single characteristic of the image retrieval can not always bring enough distinction informations, we propose to conduct the retrieval by the new method of combining more characteristics to get the best effect. The result of experiment of the trademark image retrieval system indicates that the new method works better than the retrieval based on the single characteristic. Since the user of the image retrieval system is human being, it is very important for us to get a comprehensive understanding of image content through the interactive way. So it is necessary to add some text information in the system, namely to conduct the retrieval based on the image understanding of users. The experiment result shows that the effect of retrieval based on both text and content conforms to the human visual sense.