Improving the Accuracy of License Plate Detection and Recognition in General Unconstrained Scenarios

Nowadays, Automatic License Plate Recognition (ALPR) is widely used in commercial applications. Some of the existing methods focus on the detection of license plate (LP) images with approximate frontage, while others deal with the unconstrained LP of images, but the result of LP recognition under some extreme conditions is unsatisfactory. This work mainly focuses on improving the existing unconstrained LP capturing method [1] to extract the LP and recognize it in several extreme cases as many LPs may be seriously distorted, even not parallelogram. Our main contributions are the following. First is to modify the loss function of the existing Convolutional Neural Networks (CNNs) so that the output parameters can form a parallelogram directly, then, LPs could be detected more accurate in a single input image first. Secondly, two parameters need to be estimated are added to the output of the networks' fully connected layer structure to make the corresponding output parameters form an arbitrary quadrilateral, which is closer to the LP shape of the actual imaging. In this way, it can detect and correct LPs with various shapes (LP may be deformed into irregular quadrilateral due to various reasons) in a single image. Thirdly, a feedback mechanism would be added to identify and process the LP images which are not detected correctly after the LP characters are recognized by Optical Character Recognition (OCR) network, then, return the processed image to the detection network for re-detection. In general, LP can be detected correctly. Compared to the other benchmark methods, the experimental results have demonstrated that the proposed method achieved the best performance, especially in the extreme conditions.