Novel System for Retrieval of Composite Trademarks Using Multi-feature Voting

Trademark retrieval systems have been a well researched field however majority of these researches have been done on device trademarks and do not consider the presence of text embedded within trademark images as in case of composite marks. In this work a unified retrieval system has been proposed and implemented for composite trademarks. The technique is invariant to font size, font style and orientations, further it can even apply to non standard/ artistic fonts. Three recognition techniques have been proposed based on directional chain codes, curvature of boundary and distance from centroids and a voting algorithm has been proposed to combine their results so as to avoid false positives in the result. Evaluation of the technique has been done on one benchmark dataset and on the self compiled dataset. The results obtained establish the proposed technique as a viable solution and effective contribution in the field of composite trademark retrieval. Owing to the significant advantages associated with the proposed method, an improvement of 13.5% in recognition accuracy is achieved.