Repository files navigation

An experimental implementation of MultiRes-U-Net for facial feature segmentation using the CelebAMaskHQ dataset.

The aim is to utilise and discover the compatibility provided byMultiResUNet models, originally utilised for Medical Image segmentation, and check thefeasibility of utilising the aforementioned method in order to provide segmentation of facialfeatures (lips and eyes in our case).

Refer thepdf for further details on the documentation, code implemented and directions to follow before running the notebooks.

NOTE: The original implementation was created onGoogle Colab, and can be testedhere with the sample implementations provided. I would recommend colab for viewing and using these notebooks for any and all purposes as many "improvements" are made in order to compensate for restraints pertaining to both Google Drive and Google Colab (eg. The workaround notebooks for time and RAM restraints in the /WorkAroundNB section of this repo) which are visible in the notebook code and comments.

Link to theCelebAMask-HQ dataset used for this investigation

Link to the original MultiRes-U-Net implementation ishere.

Output for a random image using the FD10 model in the project.