Abstract:
In this dissertation the problem of recognising vehicle licence plates of which the sym¬bols can not be segmented by standard image processing techniques is addressed. Most licence plate recognition systems proposed in the literature do not compensate for dis¬torted, obscured and damaged licence plates. We implemented a novel system which uses a neural network/ hidden Markov model hybrid for licence plate recognition. We implemented a region growing algorithm, which was shown to work well when used to extract the licence plate from a vehicle image. Our vertical edges algorithm was not as successful. We also used the region growing algorithm to separate the symbols in the licence plate. Where the region growing algorithm failed, possible symbol borders were identified by calculating local minima of a vertical projection of the region. A multilayer perceptron neural network was used to estimate symbol probabilities of all the possible symbols in the region. The licence plate symbols were the inputs of the neural network, and were scaled to a constant size. We found that 7 x 12 gave the best character recognition rate. Out of 2117 licence plate symbols we achieved a symbol recognition rate of 99.53%. By using the vertical projection of a licence plate image, we were able to separate the licence plate symbols out of images for which the region growing algorithm failed. Legal licence plate sequences were used to construct a hidden Markov model contain¬ing all allowed symbol orderings. By adapting the Viterbi algorithm with sequencing constraints, the most likely licence plate symbol sequences were calculated, along with a confidence measure. The confidence measure enabled us to use more than one licence plate and symbol segmentation technique. Our recognition rate increased dramatically when we com¬bined the different techniques. The results obtained showed that the system developed worked well, and achieved a licence plate recognition rate of 93.7%.
