Research On Image Repair Method Based On Generative Adversarial Network

Image restoration is a relatively tedious and delicate work,and its technology has a wide range of uses in the protection of cultural relics,film and television production,and criminal investigation.The use of deep learning related knowledge for image repair can not only ensure its repair effect,but also has important practical application value.This thesis mainly conducts in-depth research on face repair methods based on generative adversarial networks,and proposes an improved repair method based on it.Since most of the repair algorithms use the entire image for repair or artificially determine the area to be repaired before repair,the repair work is cumbersome and the repair effect is general,so this thesis introduces the image segmentation algorithm.In addition,based on the defects of the existing face repair algorithm for generating adversarial networks,this thesis proposes to change the ordinary convolution into a deformed convolution and add a spatial pyramid technology to solve the defects of the original model.The experimental results show that compared with the original model,the improved model not only has a significant improvement in repair effect,but also improves its repair efficiency.The main work of this thesis is divided into three parts:1.Focuses on the use of image segmentation methods.Among them,image segmentation mainly uses Robert and Canny combination of two basic operators to segment the image to obtain the image to be repaired area.The main reason for using two operators is that a basic operator is used alone,and its segmentation is not only unstable,but also the segmentation effect is not ideal,and the combination of the two operators can improve the segmentation stability and segmentation effect.In addition,the complex segmentation algorithm is not used,mainly to ensure its time consumption.Experimental results show that the segmentation effect is better and the time taken is shorter.2.Improve the generator part and discriminator part based on the original generation confrontation network.First,based on the original network,it can only process specific images and models to obtain the instability defects of the spatial features between pixels in the image.In the generator part,it is proposed to replace the deformed convolution with the ordinary convolution and add a spatial pyramid to the pooling layer.technology.In the discriminator part,because the segmentation results obtained by the segmentation algorithm are inconsistent,two different discriminant methods are used.One of them is local image generation,which is based on a good image segmentation effect.The improved method is to add a contour discrimination before the original local and overall discriminators,which speeds up its repair to a certain extent.effectiveness.The other is the overall image generation,which is based on the unsatisfactory segmentation results.The method is to use only one overall discriminator.The main reason is that the overall image is generated directly in the generation part because of the unsatisfactory segmentation,so local discrimination is meaningless.This method is mainly to supplement the local generation and make the model more complete.The experimental results show that this method not only effectively improves the repair quality,but also improves the repair efficiency.3.A face restoration system is designed and implemented by using MATLAB GUI.This thesis is mainly based on the improved repair algorithm,and then combined with the existing system design knowledge,designed a face repair system to meet the needs of ordinary users.