Context and problematic

This project is related to the skin analysis project. Once the scores had been predicted, L’Oréal wanted to set up a visualization system (segmentation) of the areas which were used to predict the scores.

Goals

The aim is to show that the skin analysis algorithms are based on the right information present in the image.

Our intervention

2 Data Scientist in SCRUM mode

• Creating and cleaning the dataset
• Data preparation (normalization, data increase on the fly, etc.)

We have developed a solution, which is based on the different layers of the algorithm, to highlight the pixels used for the prediction.
Moreover, we completed some visualizations through a GAN changing the age to zero to recover a wrinkle differential. The results were then improved and made more aesthetic through several processes (alteration of the alpha, heat map, guided filtering…)

Results

Visualizations can be carried out in groups or individually
They are very robust and have been validated to be deployed on machines in the shop.
The operations allowing these visualizations have been successfully tested on mobile

Technical environment

Python – Pytorch – Tensorflow – Pillow – Scikit Image – Innvestigate – Onnx – Tflite – Tfjs – PytorchMobile – pandas
GAN
Docker – Conda – Jupyter
Git – GitLab
Linus – Cuda