Research On License Plate Recognition Based On Accumulated Pixel Value Comparison And Minimum Enclosing Rectangle

In recent years, as an important research direction of Intelligent TransportSystem (referred to as ITS), license plate recognition system is increasingly paidattention to by scholars. It is an application of computer video image recognitiontechnology in license plate recognition, it can be applied in road monitoring andalarm, speeding violation penalties, vehicle access control, highway tollmanagement, automatic registration of license number, etc. All that is importantfor the maintenance of law and order in the city, the improvement of roadconditions, and the achievement of automated traffic management.Through the in-depth research to the status quo at home and abroad andthe existing recognition algorithms,this paper proposed a license platerecognition algorithm based on accumulated pixel value comparison andminimum enclosing rectangle(referred to as MER). The algorithm includes threemodules of license plate location, character segmentation and characterrecognition. The major research content is listed as below:(1) An algorithm based on wavelet transform’s multi-resolution idea andMallat transform was used to detect and extracted the edge of license plate, therunning time of the image processing has been shortened. (2) The initial license plate location was achieved by prior knowledge. Inthe process of precise location, horizontal positioning was achieved by thealgorithm that compares the accumulated pixel value with set threshold from thecenter position of the matrix of horizontal projection of the image to both sides,vertical positioning was completed by the method, in which the independentareas in the image are marked with MER and the coordinate information ofMER was also combined.(3) This paper proposed an improved Otsu algorithm based on globalthreshold and reclassified the region on the two-dimensional histogram, theimage binarization was quickly completed. Tilt correction was achieved byRadon transform and coordinate transformation. Then, Character segmentationwas finished through vertical projection.(4) The character recognition algorithm of template matching and neuralnetwork were respectively studied and improved, this paper gave thecomparison of the two improved algorithms. The test proved that the gradientdescent algorithm with the momentum based on BP neural network has a higheridentify and less time consuming.(5) This paper built the test interface using Graphical User InterfacesDevelopment Environment(referred to as GUIDE) of Matlab, positioningaccuracy and recognition accuracy of220sample images were count up andanalyzed, the results proved that system has good real-time performance androbustness.