Contains two implementations of automatic image colorization.

Implementation 1: Given a training image (trained on pixels), we train a all-vs-one linear classifier on the U and V color channels for individual pixels. Then, we predict the test image's U and V color channels by feeding in the Y color channel one pixel at a time.

Implementation 2: We colorize an image given a semantic prior. The prior, or image class, is obtained by first training a NN classifier on scene recognition that outputs a probability vector for the different scenes, using cross-entropy loss as the error function. We use 3-fold cross validation to avoid overfitting the model. Then, given inputs of the image and the probability vector, we train an NN to output colorings of the image in the U and V color channels, using RMSE as the error function.

Image classes: beach, city, lowlands, cave, glacier