Research Of Trademark Retrieval Algorithm Based On Local Feature

In the modern business world, trademark as an important intangible asset of enterprises, is a symbol of enterprises reputation. Therefore, the same type of trademark should be distinguished in the logs classification and registration process. The traditional image retrieval methods use key words, which are manual text annotation of images, to retrieve images. But manual text annotation is time-consuming and large workload, and it depends on the people’s subjective judgment. Recently, the number of registered trademarks increases exponentially, so that the traditional methods fail to accomplish similarity comparison among massive mark images. Therefore, how to efficiently and accurately retrieval similar images from mass image data, is an important research direction in CBIR(content-based image retrieval) today.Image retrieval based on local feature points is a research hotspot in the field of CBIR. Combining the knowledge of machine learning, this thesis accomplishes a series of researches on trademark retrieval algorithm based on Sift feature, which mainly include the following four aspects:(1) In allusion to the problem that there are a lot of Sift features of one image, a simplified Sift features based on the clustering algorithm is proposed. The method not only reduces the number of Sift feature points for each image, but also not reduces the query accuracy. Therefore, this method reduces the time for feature matching and improves the efficiency of the trademark image retrieval.(2) When the feature vector is above10-dimension, the feature matching algorithm based on normal KD-tree reduces retrieval efficiency and is worse than traverse method. And when the sample size too much, only finds the close points but not the closest points. A new KD-tree based on PC A is proposed to overcome this problem. Experiments show that the method has higher retrieval precision.(3) Trademark retrieval method based of Sift feature may leave out the reversal and highly similar trademark images. A new algorithm based on Sift descriptor and corner feature is proposed. This algorithm uses corner feature to make up for the lack of Sift. Experiments show that this algorithm not only contains the characters of Sift descriptor like robust against noise, but also improves the ability of description of image shape. Therefore, this method has higher retrieval precision than other methods.(4) Based on VC6.0platform, combining feature extraction and feature matching proposed by this paper, we construct a retrieval framework with local features for content-based trademark image retrieval.