| With the development of deep learning technology and hardware technology,detection and recognition technology based on deep learning edge computing has become a major point of research nowadays.Traditional licence plate localisation is mainly performed using the features,textures and colours of the licence plate,whose main problems are non-targeting,high time complexity,poor robustness and window redundancy.Compared to traditional recognition algorithms,deep learningbased number plate detection algorithms have improved greatly in terms of accuracy and real-time performance.However,due to the complexity and variability of real scenes,it still faces many problems.It is ineffective in coping with large angle license plates,small target license plates,poor lighting environment and obscured license plate detection,high network complexity,low detection real-time rate,reliance on high-speed network,expensive GPU in front-end deployment,high limitation by latency,and inability to perform unrestricted scene license plate detection under edge computing.This thesis therefore investigates the work on license plate detection in unrestricted scenarios,and through lightweight optimization design,reduces the complexity of the algorithm,speeds up convergence,improves the accuracy of detection,and meets the needs of edge computing.The specific work is carried out as follows.(1)In this thesis,we propose a key point detection method Retin ALPR based on pyramid structure for the problem of low accuracy of license plate localization in unrestricted scenes,and use the attribute pyramid structure for evaluation training to enhance the extraction ability of useful features in the backbone network,and the expression ability of the model.The network model is also effective in locating license plates in unrestricted scenes.The AP metric is improved by 6% compared to Faster R-CNN as a target detection algorithm.(2)In order to enhance the self-learning capability of the model,this thesis investigates the attention mechanism in convolutional networks and designs an attention unit module based on the channel attention mechanism and spatial attention mechanism,and applies it to the network.The simplified lightweight pyramid structure proposed in this thesis improves the AP metrics by about10% on the test dataset,and the single image performance test only increases the elapsed time by about 0.03 s.The effectiveness of the approach is demonstrated through experimental deployment on a Raspberry Pi,comparing both accuracy and speed. |
