A MODULAR NEURAL NETWORK CLASSIFIER FOR THE RECOGNITION OF OCCLUDED CHARACTERS IN AUTOMATIC LICENSE PLATE READING

摘要

Occlusion is the most common reason for lowered recognition yield in free-flow license-plate reading systems. (Non-)occluded characters can readily be learned in separate neural networks but not together. Even a small proportion of occluded characters in the training set will already significantly reduce the overall recognition yield. This paper shows that a modular network can handle a realistic mixture of (non-) occluded characters with a 99.8% recognition yield per character.