Combination of CNN with GRU for Plate Recognition

License plate recognition has been a hot topic. Most of the existing license plate recognition solutions are mainly implemented through character segmentation and then recognition. However, these methods have been lacking in robustness and character segmentation has been a difficult problem to be solved perfectly. This thesis boils down the problem of character recognition to a problem of sequential learning. The convolution neural network is used for feature extraction to describe the high-level semantics of the image, and the GRU neural network is used as the sequence learning device to effectively model the internal relations of the sequence. Considering that the output sequence cannot be aligned with the input feature frame sequence, we use structured Loss. A background (Blank) category is also introduced to absorb the obfuscation of adjacent characters. The experimental training set of the paper is more than 10,000 plate data sets in the nearly real scene produced by human, and the test results of 99% accuracy can be achieved on hundreds of test sets.