Traffic Violation Image Processing Based On Deep Learning

At present,the way of capturing traffic violations in China is relatively simple,and it is difficult to distinguish the behavior of vehicles that are forced to violate regulations due to road conditions.The illegal pictures obtained by the snapshot will also contain some misjudgments.These illegal pictures require a lot of manual time for secondary screening.check.The purpose of this thesis is to replace manual screening by deep learning.By extracting information from traffic violation pictures,screening traffic violation pictures containing forced violations,and providing richer road condition information,the burden of manual screening is reduced.Assist manual processing to improve the intelligent level of traffic violation image processing.In this thesis,the YOLOv4 object detection network is used to further extract the road information in the illegal pictures,and mainly identify ten categories such as traffic police and vehicles.When the above targets appear on the road,it may indicate that the vehicle is forced to violate the regulations at the current intersection and that the vehicle does not follow the traffic signal due to the command of the traffic police.In the process of extracting traffic violation picture information,this thesis proposes methods for vehicle rolling over the line detection,stationary vehicle detection and detection of small targets in traffic lights.After synthesizing the OCR network license plate recognition results,the enhanced images processed by GAN and the YOLOv4 detection results,the traffic violation pictures containing specific categories were screened,and the screening recall and precision rates reached 93.96%and 97.71%,respectively.The license plate recognition results of vehicles with special behaviors such as turning,driving on the line,and crossing the stop line at a red light are marked to assist subsequent manual processing.The mark-recall precision rates are 73.50%and 91.56%,respectively.This thesis improves the YOLOv4 according to the characteristics of traffic violation pictures,and changes the Mosaic data augmentation method in the YOLOv4 to the Mixup data augmentation method.A variety of attention mechanisms are compared,and the Coordinate Attention mechanism is selected to join the YOLOv4 backbone.The Depthwise Over-parameterized convolutional layer is used to replace the convolutional layer in the YOLOv4 backbone,and finally the detection accuracy of the YOLOv4 network is increased from the initial 78% to 82.5%.For the edge deployment situation,this thesis makes a lightweight improvement on the YOLOv4 network,and changes the YOLOv4 backbone network to a lightweight Mobile Net V3 network.And the network is processed to further reduce the amount of network parameters,and greatly improve the network detection speed under the premise of meeting the detection requirements.Aiming at the need for vehicle license plate recognition in the process of processing traffic violation pictures,this thesis uses an Optical Character Recognition(OCR)network to recognize license plates,and associates the detected license plate results with the vehicle location.For edge deployment,an ultra-lightweight OCR network is used to recognize vehicle license plates,and the OCR network is improved according to the characteristics of the license plate recognition task.The direction classification network in the OCR network is deleted,the network structure is simplified,and the processing efficiency is improved.In view of the fact that the edges of some vehicle license plates in illegal images are not clear and there is motion blur,this thesis uses Generative Adversarial Network(GAN)to enhance specific areas of the image.The Real SR network is used to enhance low-resolution image information,improve the positioning accuracy of license plates and the accuracy of text detection.The Deblur GAN-v2 network is used to deblur some vehicles in the traffic violation pictures,so as to improve the detection rate and detection accuracy of vehicle license plates.According to the characteristics of the task,the Deblur GAN-v2 network is lightened to further speed up the network processing speed.In this thesis,the above four parts are combined together to build a traffic violation image processing system,which completes the automatic screening of traffic violation pictures in some scenarios,improves the vehicle image quality and license plate recognition effect in traffic violation pictures,and achieves better application results.