Converting gray scale images to colour using deep neural networks.
We trained and tested the model on a subset of two larger datasets. This was done to examine the performance of smaller datasetsthe colourization task. The model architecture was using an encoder-decoder model with a pretrained imagenet model feeding into the decoder.
To explore a variety of domains for colourization techniques, two different datasets were used and analyzed. The first dataset used to train the neural networks with is the CelebA dataset. This dataset is composed of a diversity of face attributes with many images of celebrities. The images in this dataset have many properties that can help neural network models find patterns such as faces that are wearing hats or glasses, faces that have facial hair, and even different expressions on faces. This dataset will examine the effectiveness of colourization on images with people.
Another dataset used to train the models is sample sample of the ImageNet dataset, called Imagenette. The original ImageNet dataset is based on the words in the WordNet hierarchy and contains over 14 million images divided accross 1000 classes. The Imagenette dataset is a lot smaller and contains only 10 classes of objects. By having different types of objects, this will test if the neural network can differentiate the colouring based on context and patterns. Overall, this dataset evaluates the colourization of more common objects, places and landscapes.
We first convert the images to grayscale and then begin training.
Ilan Gofman: @ilangofman
Daniel Truong: @daniel-truong
Vatsan Prabhu: @vatsanp