Object Recognition Based on Invariant Moments

We investigate the properties of invariant moments of binary images, it is necessary for the formation of their set in order to recognize graphic images. Pattern recognition and distance measurement used Hu invariant moments. It is shown that the invariants have different sensitivity to changes in input data that defines the strategy of their choice. Experiments on pattern recognition of text characters, images of the aircrafts and the landing site in the form of a cross were carried out. The considered recognition algorithm works in real time, use only one camera, is invariant to rotation, shift and scale the object in the frame. The accuracy and completeness of recognition amounted to about 92% on a set of thousands of samples of each type. The results of the experimental determination of the quality of recognition of various objects based on their contour images, as well as the results of comparing recognition using a different set of invariant moments are presented. It is shown that the inclusion of the less sensitive invariant moments reduces the computation time, and lowers the computational error that occurs when fluctuations in the parameters of an object or scene in the frame take place. It is proposed to combine the method of invariant moments with probabilistic neural network, which will improve the quality of recognition, making it more fast, accurate and complete.

invariants, recognition, neural network