Research And Application Of Image Filling Algorithm Based On Image Structure And Attentional

Due to the wide spread of digital images and the rapid development of computer technology,digital image processing technology has received more attention.As the main research content of digital image processing technology,image filling technology is widely used in our life.Image filling is to fill the missing areas in the image,the filled part has the same semantic information as the original image,and has a fine texture structure,so that people will not feel unnatural at first glance.Image filling technology can be divided into image outpainting and image painting.The missing region of image repair exists in the interior of the given image,which can be inferred from the neighborhood information around it to fill the missing part.The missing area in the external painting of the image exists on the outside of the given image,and the missing area is larger,which is a kind of expansion from the inside to the outside.The existing image filling algorithms already have good filling ability for low-texture images or small missing areas of the image.When the image to be filled has complex semantic information,the distance between the effective pixel and the missing region is too far,and the missing region is too large,the filling effect is not very good.The resulting region is often blurred,artifacts,and low consistency with the given image.In this thesis,aiming at the problems of existing image filling algorithms,an image inpainting algorithm and an image outpainting algorithm are proposed,and an image filling and editing system is designed and implemented.The specific contents are as follows:1.Most of the existing image outpainting algorithms are only suitable for data sets with low semantic or low texture information,and the diversity of results generated in the generation process will be greatly increased due to the existence of invalid pixels.The generated areas often have problems such as blurring and artifacts.In this thesis,a two-stage image outpainting algorithm based on edge guidance and CBAM is proposed.The first stage is used to generate a complete edge structure map,and the second stage is to fill the complete edge structure map to get a complete color image.The edge information of the image can well express the structure and potential information of the image,so that the generated part is consistent with the original semantic information.Secondly,the CBAM attention module is added to the generation network to avoid the impact of invalid pixels on the generation results.Then,in the first stage,the VAE-GAN structure is used to replace the codec to generate a complete edge image,and the feature loss function is used to replace the original loss function.In the algorithm,globally and locally discriminators are used to distinguish respectively,which makes the generated part more realistic.From the analysis of the experimental results,the average PSNR,FID and SSIM indexes of this algorithm on the face,street view and landscape data sets are 22.7961,6.8553 and 0.7061 respectively,which are better than the existing image outpainting algorithms in terms of qualitative and quantitative analysis.Finally,the algorithm is applied to fill the missing area of landscape mosaic panorama,and the experimental results show that it has a good effect.2.The underwater image has the characteristics of low contrast,inconsistent illumination and overall darkness,which leads to the blurring,chromatic aberration and deformation of the generated image.A two-stage image inpainting algorithm based on contour generation and SENet for underwater images is proposed in this thesis.The first stage is used to generate a complete contour structure map of the underwater image,and the second stage is filled on the complete contour structure map.The quality of the training data will affect the final filling effect,so this algorithm first carries on the illumination homogenization processing to the underwater image,and obtains the underwater image with high contrast and consistent illumination as the training set.Using expansion convolution in the first stage to get a larger receptive field.Then,because some underwater images have fine texture structure,this algorithm uses SENet attention mechanism to make the texture structure more realistic.The PSNR and SSIM indexes of this algorithm are better than those of other existing algorithms on underwater data sets.Finally,the algorithm is applied to fill the missing area of submarine mosaic panorama,and the experimental results show that it has a good effect.3.Based on the above two works,a prototype system of image filling and editing is designed and developed in this thesis.The system mainly includes two functions: filling the missing area of the image and editing the image.Through this system,more flexible personalized image editing can be realized,which has a certain practical value.