Research And FPGA Implementation Of Image Haze Removal Algorithm Based On Convolutional Neural Networks

Under the influence of haze weather,the image collected by imaging equipment is seriously degraded.These degraded images will affect the performance of computer vision systems,such as target monitoring system,target recognition system,etc.Therefore,in order to improve the performance of computer vision systems in haze weather,it is very important to clear the fog image.In the aspect of hardware implementation,because the development efficiency of ASIC is low,and the field programmable gate array has the characteristics of rich logic resources,high development efficiency and strong flexibility,so FPGA is generally used for the hardware implementation of digital image processing algorithm.Firstly,this paper introduces the atmospheric scattering model,which can be used to generate fog images.Then,the relevant features of traditional algorithms for image defogging are introduced.Then,the paper introduces the relevant basis of convolutional neural network,and introduces two kinds of defogging networks based on convolutional neural network,DehazeNet and AOD-Net.Due to the influence of input image size,DehazeNet algorithm can't learn the global information of the image and estimate the two parameters separately,which will produce reconstruction error.AOD-Net algorithm adds atmospheric scattering model formula in the process of training the network,and so on.An end-to-end convolution neural network defogging algorithm is proposed,which can extract the features of the input foggy image,fuse the features and After feature learning,the clear image after defogging is obtained.Finally,the algorithm is compared with other defogging algorithms subjectively and objectively.The results show that the algorithm has better defogging effect.Aiming at the defog model proposed in this paper,an image defog system based on FPGA is designed and implemented.The hardware design mainly includes SDRAM control module,SD card control module,data transmission control module and convolution neural network calculation module.Among them,SDRAM control module controls the working state of SDRAM,SD card control module controls the working state of SD card,data transmission control module controls the transmission of read-write data in SDRAM,convolution neural network module is responsible for the implementation of defogging algorithm.The Verilog language is used to describe each module,and the modelsim is used to simulate the function to verify the correctness of its logic function.Finally,after synthesis,the board level simulation is carried out,and the software and hardware output results are compared to verify the correctness of the algorithm hardware implementation.