Convolutional Neural Networks For Automatic Image Colorization
Last updated
Last updated
The paper can be found here.
Other resources: Automatic Colorization
Fully automatic image colorization
Feed-forward, two-stage architecture based on CNN that predicts U and V colour channels
SUN database for images
Quaternion Structural Similarity (QSSIM) for quantitative evaluation
We formulate image colorization as a regression problem and CNNs are used to solve the problem.
Similarly to Larsson et al. paper, semantic information and localization are important to this problem. Hence, this paper also uses VGG-16 since it is a pretrained model incorporates huge amounts of semantic information as it was trained on a dataset with >1 million images.
Input: 224x224 RGB. Our image is grayscale though, so we need to concatenate the image one after the other 3 times.
Layers in VGG-16 that were used:
All layers before the 3rd max pooling layer.
Two more convolutional layers are output from the merged VGG-16 layers (no max pooling layers since we want the output image to have the same size as the input image)
The CNN that we're actually training:
Optimizer: Stochastic Gradient
learning rate of 8×10−2
weight decay of 8×10−7
step decay of 0.5 for every 300 epochs.
Loss Function: Since Euclidean loss was too saturated they used cross-entropy
Training Dataset: ILSVRC 2012 classification dataset
500.000 images selected randomly from database then converted the training color images into greyscale images
For each image, the U and V channels were calculated
Calculating UV:
These layers were merged and upscales to the same size as the input: 224x224x451 matrix denoted with . Shown in the figure above, is then converted into UV channels.
These two layers are concatenated, now we have a new 224x224x144 matrix called .
Input: Matrix
Target: (predicted UV values)
is width of input image
is height of input image
