I have implemented the CNN based image colorization technique as suggested by Satoshi, Edgar and Hiroshi in their 2016 Paper named as “Let there be Color! : Joint End-to-end Learning of Global and Local Image Priors for Automatic Image Colorization with Simultaneous Classification [1] (Satoshi Iizuka, 2016)”. I have tried to analyze its result on various datasets.
I have used the same model architecture as suggested in the original paper with a Low-Level features extraction layer whose output is send to Mid-Level feature extraction and Global feature extraction outputs from both are than fused together in a fusion layer and then sent for deconvolution(In original paper it Up Sampling layer was used but I have used 2D deconvolution.).
1. CelebA dataset(Large-Scale Celeb Faces Attributes Dataset (Liu, 2015) [2] )
Dataset Contains more than 200k images of celebrity faces from different countries. I used 1000 images from them as training data 200 as validation data and 200 as test data.
2. Linnaeus 5 dataset (Chaladze, 2017) [3]
Dataset contains two parts Test and Train, Train contains 6000 images divide equally between 5 classes and Test contains 2000 images divided equally among 5 classes. I used images of ‘flower’ class from the dataset because they were more colorful and thus giving model more colors to learn
[1] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. "Let there be Color!: Joint End-to-end Learning of Global and Local Image Priors for Automatic Image Colorization with Simultaneous Classification". ACM Transaction on Graphics (Proc. of SIGGRAPH), 35(4):110, 2016. link
[2] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in thewild. InProceedings of International Conference on Computer Vision (ICCV), December2015. Link
[3] Chaladze, G. Kalatozishvili L. 2017. Linnaeus 5 Dataset for Machine Learning Link