Image inpainting method based on structure and texture hierarchical prediction

Abstract

The invention discloses an image inpainting method based on structure and texture hierarchical prediction. The method comprises a network model training part and an image inpainting part. The networkmodel training part comprises the steps of pre-processing a training data set, extracting an edge structure chart, constructing and training a structure completion network N1, and constructing and training a texture conversion network N2; and the image inpainting part comprises the steps of inputting a tested image to be repaired and pre-processing, extracting the edge structure chart, using the structure completion network N1 to generate an edge structure of a defect area, using the texture conversion network N2 to generate image content of the defect area, and enabling the generated image content of the defect area to fill the image to be repaired. The method is capable of decomposing a problem of the image inpainting to problems of the structure and texture hierarchical prediction, automatically generating a defect structure, and using a repaired structure edge graph to constrain a texture generating process, thereby effectively avoiding the texture confusion and the shape distortion, and greatly improving the repairing capacity in allusion to a natural image large area defect problem.

Description

technical field [0001] The invention relates to an image restoration method for large area defects, in particular to an image restoration method based on structure and texture hierarchical prediction. Background technique [0002] Since Bertalmio M. first proposed it at the Siggraph conference in 2000, image restoration technology has been widely used in the fields of cultural relics protection, movie special effects, image lossy compression, and image / video real-time transmission. This technology automatically fills the defect area or removes a specific target based on undamaged image information, but since the image of the defect area is unknown, especially when it involves large-area image area defects, there are still many technical problems to be solved. [0003] Image inpainting algorithms based on traditional methods are mainly divided into two directions: methods based on information diffusion and methods based on sample matching. The former adopts the idea of ​​spr...

AI Technical Summary

Problems solved by technology

However, when such methods are directly applied to restoration of natural scene images containing complex structures and textures, the generated content often has problems of texture confusion and shape distortion. It can be seen that this type of method cannot handle complex when trying to understand the content of the entire image. surroundings

Images