Video Dehazing Algorithm And FPGA Implementation Based On Heterogeneous Atmospheric Light Prior

Fog is an atmospheric phenomenon.Various particles in the atmosphere(dust,water droplets and smoke,etc.)absorb and scatter atmospheric light,thus seriously reducing the image quality.Many computer vision applications of remote sensing,intelligent driving assistance systems and surveillance systems require clear input images to work properly.At the same time,with the rise of short video,more and more data are transmitted through video stream,so the demand for real-time processing of video images is more and more intense.After investigating a large number of existing image dehazing algorithms and commonly used hardware acceleration schemes,the HAP algorithm based on heterogeneous atmospheric scattering model was hardware-accelerated,and a video image dehazing system was built based on Zynq platform.Through experimental analysis,it is proved that the system has the advantages of high performance,low delay and low power consumption.The specific work can be summarized as follows:(1)The principle of image degradation in foggy days and the atmospheric scattering model are analyzed.The implementation principle and advantages and disadvantages of the dark channel prior dehazing algorithm based on the atmospheric scattering model are summarized,especially the problem of lack of real-time performance.Then the non-uniform atmospheric scattering model and HAP dehazing algorithm based on the model are analyzed.In order to meet the real-time demand of video dehazing,a hardware-accelerated IP core of HAP algorithm was designed,and modules such as side window filtering,directional filtering,non-uniform atmospheric light,and dark channel estimation were implemented.The parallelism of the algorithm was maximized by optimizing instructions.(2)Based on Zynq platform,a prototype system of software and hardware co-design.The image defogging algorithm is implemented in the programmable logic(PL)with hardware acceleration,while the video image frame extraction and display modules are placed in the processor system(PS).In order to make the system can be used outdoors,a customized Linux kernel and file system are compiled,and the prototype system is configured as an SD card to boot.(3)In order to make the comparative experimental platform consistent,according to the above design process to achieve DCP,CEP,FAST algorithm.We use the typical data sets provided by I-HAZE,O-HAZE,RESIDE to test the four algorithms,and analyze the effectiveness,performance,and power consumption of the four hardware accelerated algorithms.The experimental results show that the HAP algorithm only needs 1.7ms to process a 550*412 image after hardware acceleration,which fully meets the requirements of realtime video image processing.Meanwhile,the scalability of the prototype system is also demonstrated.