| The Intelligent Transportation System（ITS）is a traffic service system,which can greatly relieve the traffic pressure and reduce the risk of accidents.The Automatic License Plate Recognition（ALPR）,as an important part of ITS,has been widely used in traffic management,road traffic detection and automatic toll collection.The ALPR usually consists of three sub-tasks: vehicle position detection,license plate position detection and license plate character recognition.The vehicle position detection is an optional sub-task.The license plate is detected after obtaining the vehicle position,which can effectively improve the detection and recognition accuracy of the license plate.The current ALPR system mainly uses traditional image processing algorithms and deep learning algorithms to recognize license plates.The ALPR system based on traditional image processing algorithm has poor anti-interference ability,which is generally used in fixed scenes such as parking lots and community access control.The ALPR system based on deep learning has strong robustness,and can effectively overcome environmental interference,which is generally used for license plate recognition in natural scenes.When the ALPR system is applied to a scene where there is rapid relative motion between the camera and the vehicle,motion blur is often present in the image captured by the camera,which seriously affects the recognition accuracy of the license plate.The current license plate recognition methods have achieved certain research results.But these methods are difficult to efficiently remove the possible motion blur in the image,and there are still problems of low recognition accuracy and slow processing speed of blur license plates.In view of the above problems,this thesis takes natural traffic as the scene,and proposes a blur license plate recognition algorithm based on deep learning,which consists of the vehicle deblurring and detection algorithm and the license plate detection and recognition algorithm.The research work and contributions of this thesis are as follows:（1）In view of the low recognition accuracy of blur license plate in existing license plate recognition algorithms,the Vehicle Deblurring and Detection algorithm（VDD）is proposed.The VDD algorithm uses the involution convolution operator to improve the backbone network of the D2 Net deblurring detection algorithm,which effectively improves the deblurring detection ability and processing speed of the D2 Net algorithm.The algorithm can not only accurately detect the vehicle position in the image,but also effectively remove the possible motion blur in the image,which can effectively improve the detection and recognition accuracy of blur license plates.Experiments verify the effectiveness of the VDD algorithm.（2）In view of the slow processing speed of the existing license plate detection and recognition algorithms,the Light License Plate Detection and Recognition algorithm（Li-LPDR）is proposed.The Li-LPDR algorithm uses the light Ghost Net as the backbone network,and effectively integrates the YOLOv4 target detection algorithm and the Transformer feature encoder.The algorithm can use the integrated network structure to detect the license plate position and recognize the license plate characters.The algorithm does not need to perform character segmentation and position correction on the license plate,and can accurately identify the license plate characters of different lengths and regions in the vehicle image output by the VDD algorithm.The algorithm occupies less memory,has high computational efficiency,and has strong anti-interference ability.Experiments verify the effectiveness of the VDD algorithm.The proposed algorithm is validated on the CCPD public license plate dataset.Experimental results show: The blur license plate recognition algorithm composed of the VDD algorithm and the Li-LPDR algorithm improves the recognition accuracy and processing speed of the blur license plate.The research work of this thesis has great reference value for promoting the development of blur license plate recognition algorithm research in the future. |
