A comparison of two different proposed feature sets for trademark recognition using neural network

The problem of trademark recognition relates to pattern recognition. Pattern recognition needs as basis knowledge about the object. The knowledge of object can be obtained by feature extraction with image processing tools. The success of any such practical system depends critically upon how far a set of appropriate numerical attributes or features can be extracted from the object of interest for the purpose of matching or recognition. In this paper, two different combinations of image features are proposed and their comparative results for trademarks recognition are discussed. The proposed features involve some simple ratios of the image pixels as well as some geometric moments. These features are invariant to translation, rotation and scaling. The goal was achieved by segmenting the image using a connected-component (nearest neighbours) algorithm. Then the features are used as inputs for a backpropagation neural network for the learning and matching tasks. The effectiveness of the proposed feature sets is tested with various trademarks, and the results are encouraging.