Kohonen Neural Network Performance in License Plate Number Identification

Abstract

This paper presents character recognition application development and its performance for vehicle identification. The recognition employs Kohonen self-organising MAP (SOM) algorithm to recognise license plate number. Several stages have been performed in the development process. Preprocessing stage includes image normalization, image segmentation, image enhancement, and binary segmentation are accomplished using MATLAB tool. The next stage is the unsupervised neural network architecture determination and finally the development of interface, training, and testing. Characters of license plate number are captured with 640 by 480 pixels wide snapshots from the selected samples. Kohonen SOM used Euclidean distance to determine best-matching unit and utilised two-dimensional layer map. Experiments are then performed to determine radius and learning rate based on size of the map. The result has shown that 16 x 16 size of map gives better performance with 78.57 % of accuracy in recognising the particular plate number.