Trademark is not only important to identify the goods, but also represents the quality of goods and the reputation of corporate. The design of trademark also includes a lot of social and cultural connotation. In order to prevent the registration of similar trademark, we must find an effective method to find out these similar trademark images, to avoid some similar images of well-known trademarks are registered in bad faith and protect the legitimate interests of enterprises.The current traditional commercial databases are based on text-based matching algorithm to retrieval information. On the one hand, the workload of describe and mark these images by hand is enormous, On the other hand, image usually have rich content, different people have different understanding and perceptions. The annotation of image is subjective and imprecise; it will cause the next step mismatch. Therefore, the method which can quickly and accurately to find the image is urgently need.First, we unified images by pre-treatment, based on image shape features include the "main line" and "main arc"(more than a certain length of the set of line segments and curves segments of an image) to retrieval trademark image. The improved Hough transform, overcome the traditional Hough transform the shortcomingas like can not determined the endpoint of line and bad data casued by non-discrimination connect the stright lines.The improved Hough transform can determine the endpoint of the segment so we are able to get all the length of line segments and curve segments of image.Secondly, we based on image texture features to retrieve similar image. Gabor transform is an important method of texture characterization, Gabor transform is a simulation by the human visual system, so there will have a good human visual effect.Finally, this paper clusters multiple features of image by improved K-means algorithm and as different characteristics have different recognition, so we use the weight of adaptive self-regulation allocation strategy, which can legitimately utilizes all the various features. Comparative experiments show that the multi-feature fusion method is better than shape feature based image retrieval method and texture feature based image retrieval method. |