Inpainting.pptx

Image Inpainting 山东大学 胡进昌 Old Photo Restoration Some examples Text Removal Object Removal Image ：X（n pixels） Observed image：Y（m pixels belong to origin image） Mask ：A（0 or 1，0: n-m） X * A = Y = = * * Missing pixels 70%missing 50%missing Total Variation TV：（1992） Energy Minimization improve：estimate X and A Fidelity item Isotropic TV function 正则化项，L0或L1 Blind inpainting L0-norm对应矩阵中非0的个数，最小化 就意味着减少未被观察到的像素的个数。从而优化 A origin 70%missing ours KSVD 50%missing ours KSVD AOP和 [13] Impulse noise removal In the case of impulse noise, the currupted pixels are no longer equal to 0. Logarithmic transforming, not equal to –K. Use L1-norm. Better ISNR salt and pepper noise origin 0%noises ours origin 50%noise ours Random valued impulse noise Fast , good ISNR Non-Gaussian noise Using the appropriate data fidelity term, inpaint image with non-gaussian noise. Rayleigh noise Possion noise 保真项 origin poisson，50%missing ours others origin Rayleigh，50%missing ours other Rayleigh，70%missing ours