This project is about to work with UNet and removing shadow from faces.
Frontal and aligned images with a size of 160×160 from 10 individuals, each captured under 64 different lighting conditions, have been provided from the YaleB database.
According to the lighting angle, these images are classified into 5 different subsets, and the number of images in each subset is listed in the file Index.txt.
Examples of some images from this dataset are shown in Figure 1.
The unet architecture :
The aim of this project was to give a shadowed image as the input to a neural network and provide the correct shadow-free image as the output, so that the network could learn the relationship between the two. Then, when a new shadowed image is given as input, the network attempts to reconstruct the shadow-free version.
In this project, I used the UNet neural network.
For preparing the training data:
- Each image from Subsets 3, 4, and 5 was paired with all the corresponding correct images from Subset 1, and the labels indicating which subset each image belongs to were also included. The dataset was structured as follows:
(image_path, mask_path, subset_name)
For preparing the test data:
- Each image from Subset 2 was placed in the test_dataset in the following format:
(image_test_path, subset_name)
The training and testing data were then passed into a Dataloader to prepare them for input to the neural network. The UNet model was trained using the training data.
All images were resized to a unified dimension of 512×256 and normalized. Each individual was considered as one class for UNet.