用小腦模型在FPGA上作車牌辨識

[[abstract]]II Abstract Vehicle license plate recognition is a system especially for recognizing and recording the car license plate. The whole process can be accomplished by two steps namely, the car plate catching and the character recognition. The first step, or plate catching step, applies the image processing technologies, such as color graying, image binarization, edge detection, crossing characteristics, to locate the license plate and the six characters on the plate from an image picture of a car. The second step, or character recognition step, applies the technologies such as Neural Network, Template Matching for pattern recognition to identify each character found from the first step. The technologies used by plate catching and character recognition are actually using the algorithms that need large quantity of data transfer and massive repeated mathematical calculations. Currently the plate recognition system is used mostly in big parking area or in free way turnpike. They are large PC related compute network system. If the system can be made small or even portable, then it can be more widely used in car parking or help in charging of high way car patrolling. FPGA (Field Programmable Logic Array) is one of the way to have these PC software become hardware circuits (so called hardwarization). This project is trying to organize CMAC (Cerebellar Model Articulation controller) algorithm to perform character recognition, and transfer it into hardware circuit to obtain fast processing and system minimization. CMAC is mathematical equations developed by Albus in 1975 based on Mass’s Cerebellar Cortex. There is no complicated mathematical operation involved in CMAC. So it is suitable for hardwarization. Also CMAC’s learning demonstrates a fast convergence and a good generalization capability, so mostly applied to manipulator and biped walking robot in nonlinear system. The software experiment shows a license recognition ratio of 90% and a character recognition ratio of 98.3%. The hardware experiment shows a license recognition ratio of 70% and a character recognition ratio 92.3%. Gate count of chip design are 11,420.