Depth Estimation and Image Restoration by Deep Learning from Defocused Images

Saqib Nazir; Lorenzo Vaquero; Manuel Mucientes; Víctor M. Brea; Daniela Coltuc

doi:10.1109/TCI.2023.3288335

Depth Estimation and Image Restoration by Deep Learning from Defocused Images

Saqib Nazir, Lorenzo Vaquero, Manuel Mucientes, Víctor M. Brea, Daniela Coltuc

IEEE Transactions on Computational Imaging

2HDED:NET jointly estimates depth and restores all-in-focus images from defocused input through a shared encoder and parallel task branches.

PDF Code

Monocular depth estimation and image deblurring are fundamental computer vision tasks, but solving either from a single image is ill-posed. When using defocused images, depth estimation and recovery of an all-in-focus image become related through defocus physics, yet most existing models treat them separately or solve them sequentially. We propose 2HDED:NET, a two-headed depth estimation and deblurring network that solves depth estimation and image deblurring in parallel. The model extends a conventional depth-from-defocus network with a deblurring branch that shares the same encoder as the depth branch. Experiments on indoor and outdoor benchmarks, NYU-v2 and Make3D, demonstrate superior or close performance to state-of-the-art models for both depth estimation and image deblurring.

@article{nazir2023depth,
  author  = {Saqib Nazir and
             Lorenzo Vaquero and
             Manuel Mucientes and
             V{\'{\i}}ctor M. Brea and
             Daniela Coltuc},
  title   = {Depth Estimation and Image Restoration by Deep Learning from
             Defocused Images},
  journal = {{IEEE} Trans. Comput. Imaging},
  volume  = {9},
  pages   = {607-619},
  year    = {2023},
  doi     = {10.1109/TCI.2023.3288335}
}