Optimization Of Real-time Video Fog Removal Algorithms Based On Physical Model And Implementation On FPGA

At present,most of the research focuses on improving the effect of the algorithm in the field of the defogging,but it is difficult to achieve both the algorithm running time and the algorithm dehazing effect.With the increasing performance of the FPGA chip and the decreasing power consumption level,the delay-sensitive and powersensitive embedded systems are becoming more and more popular.In this paper,The main work is to optimize the defogging algorithm based on physical model and the implementation of defogging prototype system.Firstly,In this paper,we make a trade-off between the running time and the defogging effect of the algorithm.we guarantee the accuracy of defogging algorithm,the optimization of algorithm makes it meet the real-time requirements of embedded video defogging system.According to the consistency between the changing trend of ambient light and the changing trend of scene depth,and the negative correlation between transmittance and scene depth,we replace the complex soft matting method in dark channel defogging algorithm with simple mean filtering method,roughly estimates the global atmospheric light and ambient light by means of mean filtering operation on dark channel image,and then uses the method of downsampling to reduce The calculation of global atmospheric light value.The experimental results show that compared with the dark channel algorithm and the guided filtering algorithm,the SSIM evaluation index is reduced by 5% and 2%,the PSNR evaluation index is reduced by 9% and 4%,and the running speed of the algorithm is increased by 1000 times and 83 times,respectively.Although the improvement of our work has some loss on the accuracy of the algorithm,the optimized algorithm is more suitable for embedded real-time systems with delay-sensitive.Secondly,we design and implement a real-time video defogging prototype system based on the optimized physical model defogging algorithm.we analyze the correlation and parallelism of each step base on the optimized physical model defogging algorithm point at the heterogeneous parallel characteristics of the FPGA.The floating-point operation in the algorithm is converted to fixed-point operation,and the arithmetic circuits of each functional module are designed,including light and dark channel module,average module,mean filter module,global atmospheric light module,ambient light module and image recovery module.The brightness and darkness channel algorithm circuit can calculate the maximum and minimum RGB components of each pixel in a clock cycle.According to the feature that the convolution core of the mean filter is 1,the average filter module is optimized to reduce the number of D flip-flop arrays from r*r to 2*r and the number of additive trees from r+1 to 2.The division circuit in the average module is converted into shift circuit and addition circuit,and the accuracy error range is between ±1.This not only saves a lot of resources of the FPGA,but also does not affect the pipeline structure of the whole algorithm circuit,and improves the operation speed of the algorithm circuit.The table tennis operation circuit is designed to improve the data bandwidth of the video fog removal system.The experimental results show that although the processing results of the video defogging prototype system based on fixed-point operation of FPGA are 9% and 8% lower than those of floating-point operation of CPU,the processing speed of FPGA is 6 times faster than that of CPU.In this paper,a color image of 1024 *768 in size is processed in a FPGA defogging scheme,which takes 33.15728 milliseconds and has a resolution of 640 *480.The frame rate is more than 40 fps,and the power consumption of the system is about 2.015W(W).The pixel processing capacity per unit power consumption per unit time is 94.8 times that of the CPU fog removal scheme and 4.6 times that of the GPU fog removal scheme.