| Automatic license plate detection and recognition system plays an important role in intelligent transportation.It has a variety of potential applications,from traffic control to safety management,and actively promotes the construction of smart cities.At present,automatic license plate detection and recognition algorithm is a mature but imperfect technology,which performs well in specific scenes.But it is still not accurate and robust enough in complex scenes,leading to limited application scope.Therefore,this thesis makes an in-depth study on the license plate detection and recognition algorithm,and designs the license plate detection and recognition algorithm suitable for complex scenes.The main research contents of this thesis are as follows:(1)Research on License Plate Detection Algorithm Based on Edge-Guided Sparse Attention Mechanism.Aiming at the large number of interference factors affecting the precision of license plate detection in complex scenes,this thesis proposes a novel Edge-Guided Sparse Attention(EGSA)mechanism from the perspective of image filtering and sparse reconstruction.The EGSA mechanism can suppress noise and pay attention to the distinguishing feature region and important edge characteristics of the license plate,so that the model can locate more accurately in complex scenes.Aiming at the problem that the Anchor-based detection methods may predict many low-quality boxes at the positions deviating from the center of target,a simple and effective method called Water Ripple Loss Mask(WRLM)was designed to suppress those low-quality predicted boxes and improve the overall performance.The license plate detection algorithm designed in this thesis achieves the most advanced performance on the largest and most diverse Chinese City Parking Dataset(CCPD)and Application-Oriented License Plate Recognition(AOLP)dataset.(2)Research on License Plate Recognition Algorithm Based on Gated Linear Unit.Aiming at the characteristics of Chinese license plate numbers with standard customization rules,which can be regarded as a language grammar rules,this thesis proposes a license plate recognition algorithm based on Gated Linear Unit(GLU)from the perspective of language modeling.The language modeling module was constructed by using the gated linear unit of parallelization calculation combined with wide convolution to learn the grammar of license plate numbers and get the structured information.By adding language constraints,higher recognition accuracy rate is achieved on CCPD containing various complex scenes.In addition,the parallelized model is also very suitable for modern hardware devices.(3)Design of Automatic License Plate Detection and Recognition System.Based on the design goal of automatic license plate detection and recognition system,this thesis integrates the designed license plate detection algorithm,license plate recognition algorithm and interactive interface into a system.In addition,in order to further improve the performance of the system,that is,to greatly reduce the amount of model parameters under the premise of ensuring the accuracy,this thesis designs a lightweight license plate recognition model.The automatic license plate detection and recognition algorithm designed in this thesis is suitable for complex scenes,which makes the system have a wider range of application scenes and practical value. |
