TW226454B - Pattern recognition method for car plate - Google Patents

Abstract

One type of car plate automatic recognition method consists of thefollowing procedures: 1. finding car plate position After the first-order difference operation is executed on the horizontalsatial frequency of car plate image, the possible clustering segmentationof plate position are linked by the group logic relation. The correspondingarea image of each group is processed with binary threshold, then combiningthe width and continuity of character stroke as the decision making base tofind the plate position. 2. segmenting the edge of plate (a). dividing the plate image into left and right area, executing thefirst-order difference operation and profile on the horizontal direction.The results of vertical first-order difference operation are processed withtriple threshold into three conditions in order to find the up edge(Ul, Ur)and down edge(Dl, Dr) of plate strings. (b). storing the vertical profile of plate area image in X array, thenstoring the results of binary threshold after horizontal differenceoperation in Dx array. After finding the most left and right three peakpositions, the optimal left and right position of plate strings can be gotby fuzzy theory. 3. plate character segmentation The horizontal gray level wave form between the up and down edge of platestrings is processed with adaptive binary threshold, then calculating thecounts of pixels with 0 gray value in each vertical line. The vertical linewith zero count is the likely position of boundary between the characters.By referring the average distance between the neiboughering characters, theoptimal character boundary can be adjusted by splitting and merging properly. 4. plate character recognition Inputting the segmented image array of each character into one neuralnetwork of feature fusion medium value operation, operating by thefollowing processing layer: (a). input layer: normalizing each character image to 16 x 16 image arraywith pixel's gray value as input data (b). input hidden layer: dividing into feature mapping layer and smoothmapping layer with inputted 3 x 3 pixel array as unit. The former is gotby backward propagation operation of neural network and the latter is gotfrom the convolution operation with smooth mask. (c). medium value operation network: processing the medium value operationwith the neural cell corresponding to the two mapping layers (d). recognition input layer: the output of medium value operation networkas inputs (e). recognition hidden layer: fully connected with recognition input layerand the value of neural cell is got by backward propagation operation ofneural network (f). recognition output layer: fully connected with recognition hiddenlayer and outputting the results of character recognition by backwardpropagation operation of neural network