| In recent years,With the development of deep learning and the innovation of machine vision technology,image defogging algorithm based on generative countermeasure network is gradually becoming the mainstream method of image defogging.Generative Adversarial Networks require lots of computational resources and higher storage costs compared with Convolutional Neural Network,Compression methods for convolutional neural networks have been proposed,but most methods focus on visual recognition rather than image generation.In order to use smaller-scale network to obtain higher-quality dehazed images,this paper starts with the structure of Generative Adversarial Networks and existing model compression methods,explores image dehazing algorithms based on Conditional Generative Adversarial Networks and model compression based on Knowledge Distillation algorithm,proposes a new image dehazing algorithm,designes the three model compression methods for Conditional Generative Adversarial Networks,and implemented the proposed dehazing algorithm on the PYNQ-Z2 hardware platform.(1)With regard to the problem of loss of high-frequency information and detail information in the classic Conditional Generative Adversarial Networks(CGAN)in dehazing,a dehazing method based on CGAN and fusion of image high and low frequency components is proposed.The haze improvement method(HLF-CGAN)obtains the high and low frequency components of all images input to the discriminator,including the dehazed image generated by the generator,the real hazefree image and the original foggy image.All images fused with frequency information are sent to the discriminator together.For the OTS and ITS foggy data sets,the final experimental data show that the dehazing evaluation indicators PSNR and SSIM of the network are increased by 8.20% and5.67% respectively,which proves that this algorithm has a good dehazing effect.(2)With regard to the problem that the Conditional Generative Adversarial Networks has many weight parameters and large scale resulting in insufficient computing power of CPU or GPU,three model compression methods are proposed,which are based on Feature Map,FSP matrix and Attention Map knowledge distillation.For the foggy image datasets,Experiments show that the dehazing performance and convergence rate of the student network based on the FSP matrix are better than the other two methods,and it is concluded that the number of parameters of the student network is only 1/2 of the teacher network,when the student network iterates the 30 kth time,the PSNR and SSIM of the generator output image are close to the values at the 50 kth iteration of the teacher network,and the training speed of the student network is improved about 1.67 times.Finally,dehazing evaluation metrics for student networks is similar to the teacher network.(3)Carry out hardware research and deployment for conditional generative adversarial networks,this paper realize the overall design of the dehazing system based on the PYNQ-Z2 hardware platform.The design of image preprocessing unit on FPGA side is realized with Vivado tool,and the design of neural network and IP core driver is realized with Jupiter Notebook on ARM side.It was finally verified that the system successfully realized the image dehazing,and the generated dehazing images also had good visual effects and evaluation indicators,which proved the accuracy of the design of the image dehazing system based on PYNQ-Z2. |
