Image-Inpainting-via-Sparse-Representation

• a project proposed by the assistant professor 'Nicolas Baskiotis' specialized in statistical / machine learning at Sorbonne university

• The main aim of the project is :

  1. to tackle a classical machine learning task in computer vision which is called " inpainting " .
  2. get in touch with papers ( especially one considered as 'the reference' ) and implement it from scratch .
  3. comparing the two famous regularizations L2 ( Ridge ) and L1 ( LASSO ) in this particular problem and fine-tuning hyper-parameters

• Inpainting is a classical computer vision problem , its goal is to fill blackholes and missing pixels of a bruited image in a way that keep the image visually realistic with a smooth texture !

Results

Inpainting :

Denoising ( even though in the papers , they said that the algorithms are not well suited for denoising we tried to adapt these latters ) :

References:

[1] Bin Shen and Wei Hu and Zhang, Yimin and Zhang, Yu-Jin, Image Inpainting via Sparse Representation Proceedings of the 2009 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP ’09)

[2] Julien Mairal Sparse coding and Dictionnary Learning for Image Analysis INRIA Visual Recognition and Machine Learning Summer School, 2010

[3] A. Criminisi, P. Perez, K. Toyama Region Filling and Object Removal by Exemplar-Based Image Inpainting IEEE Transaction on Image Processing (Vol 13-9), 2004