Research On License Plate Detection And Recognition Method In Unrestricted Scenes

License plate detection and recognition is an important research problem in the field of intelligent transportation,and in recent years,automatic license plate recognition in specific scenarios has a relatively wide range of applications.However,it is still challenging to recognize license plates in complex scenes,especially under uncertain camera views.Deep learning algorithms have high requirements on labeled data and computing power,and it is also worthwhile to study how to obtain labeled data cheaply and accelerate the inference process of algorithms so that they can be applied to practical scenarios.In view of the above problems,this thesis will research on the detection and recognition of license plates in unrestricted scenes and automatic data annotation methods,and the main work of the thesis is as follows:(1)Aiming at the problem of missing label data in license plate detection and recognition algorithm based on deep learning,a large-scale license plate dataset synthesis method based on simulation environment is proposed.Through the construction of three-dimensional virtual scene and the random combination of vehicle model,license plate,environment and illumination in the scene,automatic image synthesis and generation of annotation data can be used for detection and recognition tasks,greatly reducing the cost of manual annotation.(2)In order to deal with the license plate detection task in complex and unrestricted scenes,a corner detection model based on feature fusion is proposed.In the process of extracting target features,feature maps of different sizes are fused to improve the perception ability of the network to the license plate target.For the perspective deformation caused by the change of uncertain shooting angle,the corner coordinates are used to replace the traditional method of axis alignment bounding box to wrap the target to recover the perspective deformation of license plate.In order to make the modified detection model use less computing resources,a license plate corner detection model based on channel pruning is proposed.On the basis of the original model,a batch normalization layer is added after the convolution layer to judge the current channel value,and the useless channels for model reasoning are removed.The efficiency of model reasoning is improved without reducing the accuracy.(3)For the recognition of license plate characters,in view of the uncertain character length and the number of lines in the Chinese license plate,through the cooperation with the detection algorithm,the double line license plate is classified,corrected,segmented and spliced to make it a special single line license plate.Finally,it is effectively recognized through the indefinite length recognition model.Finally,the realization of automatic license plate recognition system is completed.The combined model can effectively detect and recognize any type of license plates in mainland China.