Research On Content Based Trademark Image Retrieval

With the rapid development of the national economy, coupled with the companies’growing awareness of intellectual property strengthened, the number of trademarkregistrations increase significantly compared to before. Trademark retrieval field has urgentneeds for the improvement of image retrieval performance, and the study of trademark imageretrieval techniques has important practical significance.As the classification and labeling process of trademark images are completed by hand,now the widely used category-based and text-based trademark image retrieval technologiesare more susceptible to human subjectivity, and the content-based image retrieval could bemore objective to reflect the essential characteristics of the image. As one of the essentialcharacteristics of the trademark, using shape features can increase the accuracy and efficiencyof trademark retrieval, and the shape of the trademark image，which is a stable characteristic，does not change with the change of the image color.After a lot of study to previous work, based on the study this paper focuses on bound andregion based trademark image retrieval method and brings Bag-Of-Features model into thetrademark retrieval field. The main work includes: based on the three feature extractionalgorithms，the improved7HU invariant moments, the eccentricity and the boundarydirection histogram, and combined with user relevance feedback, I proposed a multi-featureimage retrieval algorithm; utilize SURF (Speed Up Robust feature) feature extractionalgorithm, k-means clustering algorithm, combined with TF-IDF (Term Frequency-InverseDocument Frequency) technology to achieve the BOF model; design and implement anexperimental prototype system, experiment with7HU invariant moments, eccentricity,boundary direction histogram, integrated multi-feature algorithms, and BOF retrieval methodon the prototype system, and make an analysis of the experimental results. The retrievalresults of integrated multi-feature retrieval algorithms and BOF retrieval method are betterthan single feature retrievals’.