| With the progress of the times,the popularity of vehicles is getting higher and higher,and people's management of urban traffic is becoming more and more difficult.Intelligent transportation system is a kind of real-time,accurate and efficient traffic management system which makes full use of various advanced high-tech,and plays a very important role in improving the road capacity and traffic management efficiency.The core technology of intelligent transportation system includes license plate recognition technology,which plays a decisive role in the development speed and technical level of intelligent transportation system.License plate recognition technology is the result of comprehensive research on image processing,computer vision,pattern recognition and machine learning,which is convenient for real-time detection,monitoring and management of urban vehicles.It is to identify the corresponding license plate number from the pictures collected in the video image.Due to the complex and changeable weather in the natural scene,along with the emergence of rain,snow,fog and different lighting conditions,the collected pictures are not clear,resulting in the decline of the accuracy of license plate recognition.So how to remove the noise in the image and restore the image becomes the focus of our research.The main contents of this paper are as follows:1.In order to evaluate the quality of unreferenced image,a pseudo reference image guided quality regression network is used.The pseudo reference guided quality regression network can be used to simulate the behavior of human visual system,and evaluate the quality by using the perception difference information between the distorted image and the pseudo reference image.This method is mainly divided into two parts.Firstly,a pseudo reference image is generated based on the distorted image to make up for the lack of real reference information.Then,the difference graph between the natural coded distorted image and the pseudo reference image is obtained to guide the learning of the regression network,so as to evaluate the quality.2.The convolution neural network based on attention mechanism is used for image restoration after quality evaluation.Firstly,the context information is aggregated by fusing the feature graphs of different scales on the single path convolution structure,which reduces the information redundancy of the convolution layer and reduces the calculation parameters.Secondly,the attention mechanism is used to learn the local and global information of the multi-scale fusion feature map.The attention map is generated successively from the separation space and channel dimensions,and then the attent ion map is multiplied by the input feature map for adaptive feature extraction.Finally,using the low-level and high-level visual task cascade solution,on the basis of not changing the existing data set and license plate recognition algorithm,combining reconstruction loss and prediction loss to train the image restoration network,improve the perception quality of the restored image and the accuracy of license plate recognition.3.For the recovered image,the end-to-end multi task convolution neural network method is used for license plate recognition.First,the recovered image is detected by the detection network,and then the multi task collaborative computing method is used,that is,in the training process,the CNN network is trained simultaneously by four tasks: license plate classification,boundary box regression,license plate location and license plate color recognition.Training multiple related tasks at the same time can ensure independence and relevance,and improve learning performance.Fina lly,an end-to-end license plate recognition algorithm based on C RNN and CTC is used to directly output license plate characters without segmentation,which effectively solves the problem of the adverse impact of character segmentation on the recognition accuracy in traditional license plate recognition methods,so as to obtain high recognition accuracy. |
PDF Full Text Request