Extraction And Inpainting Of Occlusion Regions For Presentation Images

As an important part of people's work and life,presentations created by Power Point have been widely used in classes,conferences,lectures,etc.Audiences usually import presentation photos into note taking applications for review.However,limited by environmental factors or photography skills,some occlusions like speechers' bodies or platforms may interrupt the presentation regions,which affects integrity and visual aesthetics of images.Therefore,research on detection and inpainting of occlusion regions in presentation images will be a significance task.Existing image inpainting methods often lead to artifacts such as color discrepancy and blurriness when facing presentation images with large masked holes or sharp edges.To address the problem,this paper proposes a method to extract and inpaint occluded regions of presentation images.Firstly,the system uses Grabcut to segment photos into foreground part and background part,where foreground part including presentation regions,and background part including occlusions.Then contour detection and perspective transformation are applied to get the presentation regions and occlusion masks.After that,the system follows a two-stage method that comprises of an edge generator followed by color completion module.The edge generator produces edges of the occluded regions,and color completion module use multi-scale Patch Match to fill the occluded regions.Propagation initialization is used to generate placeholders for the occluded regions in smallest scale,then keep scaling up until the images meet their original scale.In each scale change,Patch Match with hallucinated edges constraints improves quality of the occluded regions and reproduces fine details.Using multi-scale Patch Match with propagation initialization to obtain the pixels of the occluded regions can effectively accelerate the convergence speed of the algorithm and enforce perceptually realistic results.A small presentation image dataset and occlusion mask dataset are established for experinments and evaluation.Experimental results show that the system achieves a higher accuracy than 89% in presentation and occluded regions segmentation.Besides,the inpainting module achieves 30.53 in PSNR and 0.96 in SSIM,better than some existing methods quantitatively and qualitatively and satisfies the practical needs.Finally,the research results can be extended to multiple fields like photo editing,vedio inpainting and document processing.