Research On Image Completion Algorithm Based On Generative Adversarial Networks And SOC Implementation

Image is an important part of information.Digital Image Inpainting technology use the neighborhood information of the image missing area,according to rules to achieve image inpainting,so as to achieve the purpose that the image cannot be perceived visually once missing.According to the size of the repair area,it can be divided into image patching problems for small areas and image completion problems for large areas.The missing area is too large and it is difficult to complete the image.Recently,CNN has performed well in image processing.Generative Adversarial Networks have also demonstrated unique generation effects on generation problems.The problem of image completion is still in the research stage,neither Conventional Algorithms nor Deep Learning Algorithms can achieve satisfactory results.Therefore,this paper takes Generative Adversarial Networks as the core,studies the performance of deep learning on the image completion problems,and conducts SOC design to further improve the algorithm performance.Firstly,this article analyzes the performance of the context encoder structure in solving the image completion problem.Secondly,it analyzes some of its problems,including processing image pixels are too small,the proportion of missing areas is too large,too complicated network structure,and the amount of data is too high.In response to this,this article has improved the image resolution,adjusted the percentage of the missing area.Based on the Deep Convolutional Generative Adversarial Networks,it designs image completion algorithm.The model includes the generative network and the discriminant network.The generation network sets the input layer,subsampled layer,feature hidden layer,upsampling layer,and output layer,and the discriminant network is implemented by convolutional layers.Trained on the ImageNet data set,it is optimized the network parameters,and obtains a better-performing network structure parameter.Finally,the functional performance test of the network is carried out under the same test set,Comparing with the results of the existing network structure.Use SSIM,PSNR and other image quality evaluation indicators to evaluate network performance.The results show that the network performs better on the image completion problems.Based on the deep convolutional Generative Adversarial Networks structure in this paper,the calculation details of the image completion algorithm are analyzed,the IP is designed,and the SOC software and hardware method is used to realize the image completion function on the ZedBoard platform.Due to the complexity and repeatability of the core part of the network operation,the parallel design will reduce power consumption.Firstly,extract the network structure parameters,and fix the floating-point number to facilitate FPGA operation.Secondly,design the convolution calculation IP core,perform block design,and use FPGA to complete the convolution operation process.Finally,the ARM part design is carried out and the realization of data control between multi-level convolution is completed.After completing the SOC design,the joint test is carried out to verify the effect of image completion.The comparison of the results shows that the SOC design has realized the above algorithm and completed the image completion task.