Image Recoloring

Image Re-coloring

Colorization by example

Reported by Ron Yano and Mariann Normatov, July 2014

Based on Colorization by Example, - R.Irony, D.Cohen-Or, and D.Lischinski, Eurographics symposium on Rendering (2005)

What is this project all about?

Usage

Our approach and solution

Results

Who are we

More information on the algorithm

Terms of use

What is this project all about?

In our project we have implemented a technique for colorizing grayscale images by example .Given the input grayscale image and reference colored image (example), we are transferring the chromatic channels from the example to the grayscale image in few stages process. The result is - providing to the grayscale image automatic micro-scribbles as a preparation to the final colorization stage based on Levin algorithm.

Usage

· Color old “black and white” pictures.

· Re-color images to different color plate.

Our approach and solution

· Quick overview

The reference image is segmented according to the different regions in the image. Each region is assigned a label and is used in the subsequent supervised classification scheme for segmentation of the input image. Features are extracted from the input and reference image. The input image regions are classified using a two step process, voting in feature space and in image space. Finally the colors with a high computed "confidence" are transferred from the reference image to the input image, and then fed as input to the optimization step.

· Detailed solution and approaches

Ø Input and Segmentation

Two images of cheetah illustrated above are chosen as the input and reference image respectively.

The reference image is converted to the luminance channel and then segmented by k-means algorithm, given the parameter k.

Each segment is associated with its own label.

Ø Feature Extraction

We compute DCT of kxk neighborhood for each pixel in the reference image, matching it to its label and create in a such way a feature space.

We also compute DCT of each pixel in the input image, for classification stage.

DCT coefficients are simple texture descriptor, which is not too sensitive to translation and rotation.

Ø Classification

We first classify a novel feature vector by Knn algorithm using major voting method. Knn alg. is applied on the previously

computed feature space. However, this approach leads to many erroneous pixel classifications. Therefore we switch using

unique technique to low-dimensional feature space and only then perform the classification by Knn.

A low-dimensional space switching approach, based on examining the differences between the vectors within the same class

and differences between vectors belonging to the different classes. This method is performed by applying two-stage PCA

on the initial feature space. The advantage are illustrated below:

Ø Image Space Voting

As we can see, still many pixels of input image are misclassified. To improve our results we consider N(p), the kxk neighborhood

around a pixel p in the input image. We replace the label of p with the dominant label in N(p). The dominant label is the label with the highest confidence value conf(p,l). The confidence of pixel p with label l is defined as : , when W denotes the weightings,

which is calculated as: . The confidence conf(p,l) is typically high in neighborhoods where all (or most) pixels are labeled l, and low on boundaries between regions, or in other difficult spots. The improvement are illustrated below: