Research On Image Haze Removal Algorithm Based On Generative Adversarial Network And Hardware Acceleration

With the rapid development of informatization,images have become an important means of human information transmission.However,bad weather such as haze usually reduces the contrast,color shift degradation,and loss of details of the images acquired by imaging equipment.These low-quality images will not only affect people’s subjective visual perception,but also make it impossible for people to better obtain and understand the information contained in the image.and it will affect the performance of many intelligent technology processing systems such as self-driving cars,security monitoring systems,satellite remote sensing aerial photography,and so on.These intelligent systems require a clear image source as the basis for subsequent work.This paper designs an image haze removal algorithm based on the generative adversarial network,and analyzes the haze removal performance of the designed network.This article first introduces and analyzes some previous algorithms in the field of haze removal,introduces the atmospheric scattering model,and essentially analyzes the reason of image degradation in foggy weather.t focuses on the analysis of the DehazeNet algorithm and the MSCNN algorithm based on the haze removal algorithm of the convolutional neural network.Aiming at the haze removal method which is dedicated to estimating the intermediate parameters to solve the clear image,it will produce the shortcomings such as error stacking,etc.,a haze removal algorithm based on the generative adversarial network is proposed.The network includes features extraction,encoding,decoding,and feature fusion,and generates haze-free images directly end-to-end,avoiding estimate any intermediate parameters.The algorithm first performs multi-scale feature extraction on the input fog images,and at the same time,the image is coded and decoded and then multi-scale fusion is performed with the extracted features,and finally a clear and haze-free image is obtained.In terms of network optimization,not only the mean square error of the generated image and the label image is used to constrain the network,but also the network is jointly optimized by minimizing the confrontation loss,adding perceptual loss,and L1 loss,so that the image can better restore color and retain detailed information in the process of clearing.Finally,it is verified that the designed haze removal algorithm has good haze removal performance under PSNR,SSIM,and MSE metrics.For Hardware Acceleration,training the proposed haze removal algorithm firstly,extract the weights,biases and other parameters generated by the training to fix them.Secondly,a hardware-Accelerated SOC system architecture is designed based on the method of software and hardware collaboration,which includes the control part on the ARM and the data calculation part on the FPGA.The entire system uses the bus based on the AXI standard for interconnection,focusing on the self-defined convolution calculation module and data transmission module in the system.Finally,the vivado software is used to simulate and verify the modules designed in the system architecture,and the board is verified on the ZedBoard board.