Motion Deblurring And Car Plate License Detection Based On Deep Learning

With the development of big data and the improvement of computing power,deep learning algorithm is becoming more and more powerful.It can learn more useful features by building a model to simulate the neural structure of human brain and experiencing massive training data,so as to improve the performance of the algorithm.In the traditional machine algorithm,most of the processing is more dependent on artificial feature extraction,which is effective but not universal for specific simple tasks.The deep learning algorithm is more popular because of its strong learning ability,adaptability and high upper limit.With the increasing number of vehicles,the traffic pressure rises suddenly,and the vehicles are always in the state of motion.Motion blur is inevitable,the generation of motion blur will affect the subsequent series of operations,and plate license detection plays an important role in the whole license plate recognition system.Therefore,around the above problems,this thesis studies two kinds of deep learning algorithms,the specific research contents are as follows:1.Firstly,the motion blur removal is studied.The vehicle is always in the state of motion,motion blur is inevitable,and motion blur is a typical image degradation problem,which seriously affects the quality of the image.How to effectively remove motion blur is a practical problem.In the traditional algorithm,it is limited to specific scene and difficult.In this thesis,a multi-scale model is proposed which is used to blind motion deblurring based on GANs.The method utilizes a new block for CNNs inspired by Res2 Net.The block increase the range of receptive fields and integrate multi-scale features at the granular level.In addition,this method combine the Perceptual loss and the MSE loss as the loss function,better structural features and detailed features are obtained.Furthermore,the Wasserstein distance is used to measure the distribution of data to improve the stability of the network.In Gopro dataset,the results obtained by our method are excellent both in evaluation index and visual effects.2.Secondly,this thesis studies car plate license detection.With the development of society,the increase of vehicles makes the traffic pressure rise suddenly,so an excellent license plate detection system will reduce this pressure.Although the current license plate detection system has been widely used,most of the traditional algorithms only detect under the reasonable and effective strong assumptions of specific environment.So a more adaptive and efficient algorithm can adapt to the real environment.Therefore,an improved license plate detection algorithm based on yolov3 is proposed in this thesis.Firstly,in order to make the algorithm used in this thesis more accurate for the number of the candidate boxes and the aspect size ratio,this thesis uses K-Means algorithm to cluster its data.Secondly,in view of the large size of various kinds of images at present,this thesis changes the original network structure of yolov3,increases the convolution layer to double the input size,and adds the multi-scale fusion method to the final prediction,which improves the prediction results.In the experiment,because there are too few data sets for license plate and few open-source data sets,this thesis adopts the data set collected by ourselves,which contains 6668 pictures with license plate.This data set contains license plates under various conditions,including tilt,too bright,too dark,rainy and snowy weather,blur,etc.The diversity of data sets ensures the stability of the proposed system in different situations.The experimental results show that the proposed algorithm is better than the original yolov3 algorithm.However,the performance of the network is not good for the detection of difficult license plates,mainly because there is less challenging license plate data in the training,so the network can only learn the location of some ordinary license plates.Therefore Faster-RCNN algorithm is adopted to detect the difficult license plates.200 challenging license plate datas are selected from the data set,and the algorithm feature extraction network is changed to resnet50 to realize the detection of difficult license plate.The algorithms proposed in this thesis are studied and improved on the network structure and training algorithm.Aiming at the problem of insufficient data,we also create our own data set and apply it to the experiment,which provides a certain idea and direction for the future algorithm research and development.