Research On Camera Lens Pollution Identification Based On Deep Learning

In recent years,the research on intelligent driving has developed rapidly,includingpedestrian detection,lane detection,traffic scene identification and other tasks,which have made remarkable progress.However,the changeable driving environment will easily lead to the pollution of the sensing system,especially the surface of the high-definition camera is easy to attach dust and dirt,thus affecting the normal operation of the intelligent driving sensing system and causing unnecessary losses.However,the influence of camera pollution on intelligent driving system has not been considered in the existing intelligent driving research.Therefore,this paper will conduct research on camera pollution identification to provide theoretical basis and scientific practical guidance for the development of intelligent driving.This paper summarizes the current research status of pollution identification at home and abroad,defines the main research contents and objectives of this paper,and then introduces the basic depth learning theory to lay a theoretical foundation for the follow-up work of this paper.First of all,a study was carried out on the problem of difficulty in obtaining contaminated data sets.Firstly,the data enhancement method based on generation countermeasure network is introduced,and then based on Cycle GAN network model,Bdd100 K data set and actual manually collected pollution data are taken as training data,and simulated pollution data are generated through the improved Cycle GAN network,respectively through subjective evaluation and objective evaluation.Secondly,the pollution identification network is designed based on the portable network Mobile Net,and 4-stage Mobile Net is obtained.Then the simulation data obtained in chapter 3are used as negative samples and Bdd100 k partial data are used as positive samples to train the network.The network test results show that the sobering 4-stage Mobile Net reasoning speed in this paper is 56% faster than the original network structure and has higher recognition accuracy.Finally,in order to verify the effectiveness of the fourth chapter of the network,the network will be tested in real vehicles.Firstly,the application process of pollution identification for real vehicles is proposed,and then multi-scene test and video test are respectively carried out on the network.The results of multi-scene tests show that the proposed network can accurately identify different levels of pollution.The results of video test show that the network proposed in this paper can effectively detect the image quality,andhas high robustness and good real-time performance.The above results prove that this paper has certain reference significance for intelligent driving pollution identification research.