How can we add realistic color to black & white images? Explore the effect of up-convolutions, weight decay, and deeper architectures.

How can we add realistic color to black & white images? Explore the effect of up-convolutions, weight decay, and deeper architectures.

This report contains the most interesting logs of the training of my neural network for the W&B colorizer challenge.

This report contains the most interesting logs of the training of my neural network for the W&B colorizer challenge.

This report contains the most interesting logs of the training of my neural network for the W&B colorizer challenge. The objective is to colorize black & white pictures of flowers in the most realistic manner.  All different models are variants of the final architecture represented below. <img src="https://github.com/borisd13/colorizer/blob/master/imgs/Architecture.png?raw=true" alt="alt" width="1200"/> </p> Please consult the technical writeup and latest version of the code on [the github page](https://github.com/borisd13/colorizer).

This report contains the most interesting logs of the training of my neural network for the W&B colorizer challenge. The objective is to colorize black & white pictures of flowers in the most realistic manner.  All different models are variants of the final architecture represented below. <img src="https://github.com/borisd13/colorizer/blob/master/imgs/Architecture.png?raw=true" alt="alt" width="1200"/> </p> Please consult the technical writeup and latest version of the code on [the github page](https://github.com/borisd13/colorizer).

The objective of this challenge is to colorize black & white pictures of flowers in the most realistic manner. This process is often applied on black and white movies or for other artistic creations. However, it is traditionnally a tedious task performed manually, frame by frame. The intent of this project is to automate the colorization of black & white pictures through a neural network. While it is probably impossible to predict exactly the correct colors due to the loss of information (a tullip for example can be of many different colors), the recognition of items within a picture (tree, leaf, type of flower, sky, insect, etc) should lead to a realistic color. Contrarily to classification, every object needs to be recognized at a pixel level, similar to the instance segmentation problem. However, instead of predicting only one class per pixel, a color needs to be predicted. At a higher level, the network needs to have an understanding of the black & white image such as: "there is a dark rose, which is then probably red (other colors would be too clear), in front of a tree, most likely brown with green leaves". Its understanding leads then to the reconstitution of the image, pixel by pixel. Check out the github writeup [here](https://github.com/borisd13/colorizer).