Research On Automatic Detection And Inpainting Of Image Damaged Region

With the rapid development of the Internet,hundreds of millions of data information are generated on the Internet every day.As a kind of information between text and video,image can let people get information intuitively and quickly from it,and it is often used as a carrier of people’s memory in real life.A large number of images are inevitably subject to some damage,such as some information damage that may occur during storage or transmission.And there is also the growing convenience of editing images due to the prevalence of image editing tools,which has led to an increasing rate of human impact in damaged images.For example,brush smear,pixel block coverage,text coverage,and mosaic are all common types of damage,and the variety of damage types can be a challenge for image inpainting.Image inpainting requires the mask that marks the damaged region of the image to be provided,and it would take a lot of time and labor to mark the damaged region manually.Therefore,it is of practical application to automatically detect and repair the damaged region of various types of damaged images.In order to complete the automatic detection and repair of damaged images,this thesis makes a preliminary study on automatic detection and image inpainting.However,since there are many types of damage,this thesis focuses on the two common types of damage detection,which are brush smearing and mosaic.The main research content is as follows:(1)To detect the brush smear region better,an improved intensity-based iterative detection method is proposed.The brush smear region is detected by dynamically filtering the pixel blocks while iterating the color intensity in combination with other features,and preserving the pixel blocks with less internal variation.The experimental results show that the accuracy of the method is around 91% on the Places2 dataset,which is partially better than the comparison method.(2)To detect the mosaic region better,a mosaic region detection method based on line expansion is proposed..Firstly,some possible candidate points are inferred from the image edge information,and then the mosaic region is gradually expanded from the candidate points with a point-line-block approach.It can avoid the possible missed detection caused by the stacking of similar mosaic blocks and also reduce the time consumption of pixel matching,which improves the detection accuracy to a certain extent.The accuracy of the method on Places2 is around 94%,which is better than the comparison method.(3)In order to complete the automatic detection and repair of image damaged region,an image damage classification and inpaint model is constructed.The classification model is optimized by combining the characteristics of damaged images and using image edge information and intensity information as fusion inputs.The model was experimented on a dataset of damaged images generated based on Places2,and the classification accuracy of the damage types was around 90% and the classification was performed well.Then an image inpainting model based on edge information is introduced.The image inpainting model based on edge information takes the form of a dual generative adversarial network,with the first half being an edge information prediction model and the second half being an image inpainting model,and a line extension strategy is used to optimize the model and to improve the ability of the model to restore the structure of the image.The model was experimented on the Places2 dataset with good repair results.