Detection And Removal Of Nearly-Cyclic Structural Occlusion

Near Cycle Structural Obstruction is very normal in nature,which becomes a key issue in computer vision,related to segmentation and inpainting.It is a great challenge to detect nearly periodic structure,because most of the current segmentation is color-based method,but not based on nearly periodic characteristic.In addition,because of large proportion of nearly periodic structure,effect of inpainting method must be perfect.This paper presents a novel approach to restore images with fence occlusions.The approach is made up of two steps: fence detection,and disocclusion restoration.In the first step or fence detection,we use superpixel to cluster image,and fit the superpixel blocks with rectangles.Through elongation of the fitted rectangles,fence pixels are collected.From those sampled pixels a color-based classifier is used to obtain the primary shape of the fence.Then,multi-RANSAC is used for fitting multi straight lines simultaneously.Moving least squares(MLS)is used for sketching the fence structure through curve fitting.Complete fence is detected by expanding the fence structure.After the removal the fence,incomplete region can be filled by low-rank method based on the corresponding block.Thus we can restore the image with fence occlution.Then we can also do some applications.For example,some other fences can be attached to the recovered images,or the detected fences attach to other images.After Virtual processing,an beautify picture with artistic effects is obtained.Experimental results show that our method can detect the entire fence,disoccluded regions were also recovered well,resulting a clean and complete image.Six associated detection segmentation are contrasted with ours.From the experimental results,both in subjective or objective evaluation system,our approach is optimal.