Research And Implementation Of Deep Learning Based Muck Truck Detection And License Plate Recognition Algorithm

Muck truck detection and license plate recognition make great value in smart transportation and other fields.With the rapid development of cities,the demand for muck trucks is increasing,which poses a huge threat to the safety of pedestrians and has an impact on the city appearance.There are challenges in real-time detection of muck trucks in real life.Uneven illumination,blur,and various angles will affect the accuracy of license plate recognition.Therefore,a method for real-time detection of muck trucks and recognition of license plates is designed to make the muck trucks intelligently managed which has great practical significance.In this paper,we proposed a method of detecting muck trucks and recognition muck truck license plates,the following work has been mainly completed:(1)An improved binary classification network based on YOLOV3 is proposed.In order to detect the location information of muck trucks and license plates in real-time monitoring,this paper proposes a binary-classification network based on the improved YOLOv3 algorithm.Two set of small-scale feature maps in YOLOv3 are used to detect muck trucks,and large-scale feature maps are used to locate license plate.The Complete Intersection over Union(CIo U)loss is substituted for the position coordinate loss in YOLOv3.The experimental results show that the improved YOLOv3 algorithm speeds up the convergence and improves the accuracy of muck truck and license plate detection.(2)A U-shaped license plate recognition network ILPRNET(Improved License Plate Recognition Net)which integrates spatial attention mechanism is proposed.In order to accurately recognize license plate characters,this paper proposes a license plate recognition network ILPRNET to ensure consistent data distribution and rapid network convergence during the training process,and enhance the network’s nonlinear expression ability as well as prevent network overfitting at the same time.Not only that,this network integrates the spatial attention mechanism to help the network pay attention to the spatial location information of the license plate character features.By introducing a symmetrical network(encoder-decoder architecture)with a U-shaped structure,each character of the license plate can be accurately located.Compared with other algorithms,our algorithm avoids the problem of low accuracy caused by inaccurate character segmentation.Experimental results show that the proposed model performs well in both ordinary and complex situations.(3)A robust license plate recognition model based on Bi-LSTM(Bi-directional Long Short-Term Memory)under non-restrictive conditions is proposedIn order to further improve the accuracy of rotating license plates recognition,this paper proposes a robust license plate recognition model based on Bi-LSTM.To fully extract the license plate features,we incorporate the structure of Bottleneck,separate convolution,and spatial attention mechanism into the network to perform shallow feature and deep feature fusion.By introducing the Bi-LSTM structure,the network can accurately locate the position of each character on the license plate.Experimental results show that this method has good robustness for regular license plates or irregular license plates,normal scenes,or complex scenes.Compared with the algorithm in Chapter 4,our proposed algorithm has higher accuracy on rotating license plates.