This paper presents a new approach for real-time license plate detection based on vehicle and text regions. Firstly, vehicle regions are extracted by single shot multibox detector (SSD) framework. Secondly, multichannel maximally stable extremal regions (MSER) algorithm is used to generate character candidates in the vehicle regions. Using properties of vehicle regions, this paper filters out false character candidates and then constructs license plate candidates with remaining character candidates. Then, false license plate candidates are eliminated by exploiting the correlation between dimension of vehicle and license plate. Finally, remaining license plate candidates are passed to a word/no-word classifier to keep the final license plate. To run in real time on embedded systems, this paper chooses the MobileNets architecture for deep CNN configurations. Experimental results on the public test dataset and new collected dataset show that the proposed approach can apply to different types of license plates with better performance than current stateof-the-art methods.
展开▼