Application of a hybrid orthogonal function system on trademark image retrieval

This paper presents a new class of discrete orthogonal moments, W-moments, based on a hybrid orthogonal function system, W-system of degree one. The W-moments have two main advantages, one is computationally simple due to the piecewise linear structure of the basis functions, and the other is the reproducibility, namely shapes and even shape groups can be precisely reconstructed by finite number of the W-moments, due to the orthogonality of the basis functions on a discrete data set. These properties make the W-moments preferable to the conventional moments when used for describing complex shapes consisting of multiple disjointed curves. To evaluate the proposed moments, image retrieval experiments are conducted on MPEG-7 CE2 and a self-constructed trademark data set. In the experiments, the W-moments are first derived from the trademark images, similarity between query image and each trademark in the database is then evaluated using the Euclidean distance between their W-moments feature vectors, and the trademark retrieval is finally accomplished by comparing their similarities. The experimental results show that the moments proposed in this paper give preferable retrieval results in comparison to other five conventional moments, and the W-moments should be feasible and efficient in practical application.