Improved tensor voting for missing edge inference
Xiangyu Hu, Ran Ma, Tong Li · 2017
Edges and structures are a critical part for natural images, which can help to improve the overall quality of the reconstructed image. In tasks like object removal, image inpainting or video error concealment, many related methods try to infer and recover the edge splines in the unknown regions using known splines in the neighbours as a pre-possessing procedure. In this paper, the tensor voting method is improved to pursue better inference quality of the missing edge. In the proposed method, the additional spline pairing is used to determine the most possible connectable edge spline pairs existed in the known region, which greatly overcomes the difficulty of choosing the accurate parameter for stick voting. Experiments show that the proposed method can improve the inference performance dramatically with the help of extra pairing work and acquire a better structure inference quality compared to several other inference methods.