Research On Image Local Features And Its Application On Trademark Retrieval System

With the development of global economy, the quantity of trademark increases dramatically year by year. Trademark retrieval has become an attractive topic in the field of image application. Retrieval efficiency and accuracy are two important issues in designing a content-based database image retrieval system.The main drawback of the existing trademark retrieval methods is their vulnerability to position, viewpoint, scale and illumination changes of image. In order to improve the efficiency and accuracy of the system with the combination of image recognition algorithms, image local feature method should be implied in the system. However, traditional image local feature also has its weakness. Hundreds of local features could be extracted from a single image. Moreover, about100elements are needed to represent one single local feature. Matching or querying with local features would result in significant computational cost.In order to reduce the computational burden as well as maintain the recognition rate of the image local features, a binary projection method for image local descriptor is proposed. The image patch is projected and transformed into a binary string for boosting the performance as well as speeding up the matching speed. The projection matrix is optimized by machine learning method to maintain its recognition rate and robustness. The experimental result indicates that only a32bits binary string is needed to perform as good as the state-of-art descriptors and it shows significantly faster matching speed. It could be applied in the trademark retrieval system appropriately.A weighted value similarity computational method is also proposed to be combined with the image local feature algorithms. The weighted values are trained with test images dataset in order to improve the system performance. The features of the query image uploaded by user will be extracted and matched with the features that stored in the image database. Then the weighted value similarity computational method is applied to calculate the similarity between the query image and target trademark. Finally, the retrieval result will be presented by the system. The experimental result indicated that the weighted value similarity computational method makes a great contribution to the system. The recognition rate of the system is improved significantly.The content based trademark image retrieval system, developed in Windows Visual Studio2010environment, is based on Browser/Server framework. In this system, many technologies are used, such as.NET Trip-Layer Framework, SQL server2008, and jQuery library scripting. The experimental result shows that the computational speed and accuracy of the retrieval system has significantly improved.