ROI-based Deep Learning Method for Variable-length License Plate Character Segmentation

License plate character segmentation plays link function between license plate detection and recognition. It is based on the results of license plate detection and produces segmented characters for subsequent recognition module. Character is the region of interest (ROI) in the license plate. Previous works focus on fixed-length license plates and face challenging on the slant plates. To solve the problem, we propose an ROI-based deep learning method for variable-length license plate character segmentation. It treats all the characters as objects and converts character segmentation as object detection. This paper exploits Faster R-CNN to detect the possible character regions. Region Proposal Network is designed to provide sufficient proposals for character detection of license plate and full connected network can modify the candidate boxes and predict the character class simultaneously. Moreover, we create a dataset for character segmentation of license plate. Experimental results demonstrate that our method can achieve a high accuracy in three scenes compared with some state-of-the-art approaches.