Research On Image Restoration Algorithm Based On Generative Adversarial Networks

When we take pictures,due to the hardware conditions of the smartphone,camera,etc.or the shooting environment,the resolution of the obtained image is too low;or the image motion is blurred due to the relative motion between the device and the target object.In both cases,low-quality images are obtained,which cannot meet the research needs in the fields of surveillance video,medical images,and remote sensing.Super-resolution reconstruction technology and image de-blurring technology can deal with the above two situations separately,restore the image and improve the image quality.The main research content of this thesis is the image restoration algorithm that based on generative adversarial network.In this paper,the 4x factor super-resolution reconstruction algorithm for low-resolution images,the de-blurring algorithm for motion blurred images,and the super-resolution restoration algorithm for lowresolution and blurred images are studied.The main research work is as follows:(1)Aiming at the problem of image degradation,we analyzed and studied the development status and characteristics of the two techniques of super-resolution reconstruction and de-blurring.The principle analysis and formula derivation of the convolutional neural network and the generative adversarial network related theory are carried out.We did the experiment of generating handwritten numbers based on the generative adversarial network.The experimental results show that the network can generate fixed handwritten numbers or arbitrary handwritten numbers,laid the foundation for subsequent image restoration studies.(2)Aiming at the problem of super-resolution reconstruction of low-resolution images,we studied the image super-resolution reconstruction algorithm based on generative adversarial network and removed the batch normalization layer in the residual module in the generative network.We used the absolute value loss function and the mean square error loss function on the Set5,Set14,BSD100,and Urban100 datasets to compare the improved network with the original network.The results show that the improved network has a training time reduction of about 18% under the same settings.The objective quality evaluation indicators PSNR,SSIM,MS-SSIM,and FSIM were used to evaluate the reconstruction results of each dataset.The evaluation results show that the improved network reconstruction is better than the original network.(3)Aiming at the problem of image motion blur removal,we studied the image deblurring algorithm based on generative adversarial network.We constructed the self-made datasets by simulating motion blur kernel and image convolution and trained the network based on the homemade dataset.The trained network and the traditional method are used to compare the blurry image processing.The experimental results show that compared with the traditional method,the method based on generative adversarial network is less time-consuming,and there is no need to estimate the fuzzy kernel,and the ringing is avoided,the recovery result is good.(4)Aiming at low-resolution motion blurred images,we proposed an image super-resolution restoration algorithm(DBSRGAN)based on generative adversarial network.The generated network consists of deblurring,feature extraction,and reconstruction modules.The two output results of the generated network are discriminated using two discriminant networks.We trained the proposed network,the experiment is carried out on the self-made blurry dataset after the training is completed,and evaluated the results from two aspects of subjective and objective indicators.The experimental results show that the proposed DBSRGAN algorithm is good.