Research On Video Denoise And Dehaze Method For Vehicle Smart Chip

The vehicle-mounted intelligent chip is the core component of the vehiclemounted intelligent system,which is used to realize the operation and control functions of the vehicle-mounted intelligent system.However,on-board smart chips inevitably need to process real-time signals of extreme noise or haze weather,which puts higher requirements on video processing algorithms.Therefore,this paper combines the advantages of traditional signal processing and emerging neural network algorithms,and also According to the requirements and limitations of the vehicle-mounted chip,a noise reduction algorithm and a defogging algorithm for the vehicle-mounted smart chip are proposed.The main research contents and innovation points are summarized as follows:(1)This paper designs a novel video denoising algorithm MVD-KB(Multi-stage Video Denoising Based on Kalman-Bilateral Filter)for vehicle scenes.First of all,traditional algorithms often regard video noise as additive Gaussian noise,but the real noise distribution is more complex.Aiming at the problem of inaccurate noise estimation,this paper studies the actual noise model in the imaging system,and proposes an adaptive noise intensity Estimation method to adapt the noise distribution in practical applications.Secondly,aiming at the "smear" phenomenon caused by timedomain filtering of moving objects in the video,by studying the theory of human visual imaging system,the new algorithm adds multi-resolution background segmentation and multi-channel noise reduction strategies.Finally,the experimental results show that even in the extreme noise environment,the algorithm can well suppress the common temporal noise in the video,eliminate the "smear" phenomenon,and the adaptive noise intensity estimation also eliminates the noise that the denoising algorithm usually has.Required pre-calibration steps,easy to deploy into chip architecture.At the same time,this paper compares the results of other mainstream algorithms,and proves that the proposed method has a good suppression effect on video noise.(2)An unsupervised learning method UDNGP(Unsupervised Dehazing Network Guided by Physical Priors)based on physical model dehazing is proposed for the real haze scene encountered in the car video.First of all,this paper designs a new unsupervised learning network based on the atmospheric light physics model,which avoids the lack of pairing of real haze datasets in supervised networks.And the problem of poor performance in real scenes caused by the use of synthetic haze dataset training,it also solves the problem of unsupervised network training,which is expensive and difficult to converge.Secondly,this paper uses the idea of multi-resolution to design a new transmission map estimation network T-Net and air light estimation network ANet,and uses Soft Matting to refine the transmission map to obtain a more accurate transmission map estimation effect.Finally,this paper makes an objective comparison in the public synthetic data set RESIDE,and the results show that the dehazing effect of UDNGP is better than the current mainstream algorithms in both SSIM(Structural Similarity)and PSNR(Peak Signal to Noise Ratio)indicators.At the same time,the subjective effect verification is also carried out on the real haze data set,and the results also show that the haze-free image restored by the algorithm in this paper is more in line with the subjective visual requirements of human beings.(3)The noise reduction and defogging algorithm proposed in this paper is integrated into the ISP(Image Signal Processor)processing module of the vehicle chip,and finally it is completely implemented on the hardware with other algorithms of the ISP module,and the vehicle chip has been taped out.In the chip test,the actual effect of the real scene proves that the algorithm proposed in this paper meets the effect requirements of the vehicle chip.