There have been times when we think about how a particular image would look in a different color space. For instance, an image taken during the day would look like at dusk.  A color transfer algorithm aims to transform the color space, mood or style of one image usually known as a reference image to another.

Applications of color transfer range from subtle post-processing on images to improve their appearance to more dramatic alterations, such as the above-mentioned daylight image to a night scene. It also finds usage in synthetic data creation or data augmentation pipelines. 

Commonly used Color transfer techniques include Histogram matching/specification and Superfast color transfer. 

Histogram matching/specification:

Histogram specification method involves spatial transformation in which the histogram of query image is matched with that of the target image. It is a three-stage process: a) Equalize the histogram of input image, b) Equalize the specified histogram, c) Relate the two equalized histograms. The mathematical details of histogram specification can be found in this referred article [4].

Superfast Color transfer:

Superfast color transfer first transforms both source and target image into the decorrelated color space (from RGB to LAB) and then uses the mean and variance statistic to transfer the luminance and chrominance component of one image to another. All the steps involved in Superfast color transfer are represented in the block diagram, depicted in Figure 1.

Figure 1  Steps involved in Superfast color transfer.

Few sample results obtained from both these methods are shown in figure 2. It is clearly evident that Histogram Specification yields superior results than Superfast color transfer. On the other hand, the histogram based manipulation considers all the pixels of the image and therefore very very time expensive, whereas the Superfast method only uses two measures (mean and variance) and therefore can be deployed in real time.

Source ImageTarget ImageColor transfer results using superfast color transferColor transfer results using histogram Specification

Figure 2: Source, Target and resultant image using superfast color transfer and histogram matching.

Applications of Color Transfer in Flixstock:

Flixstock Fashion solution solutions are built on deep learning-based architectures. Color transfer finds application in data augmentation pipelines. Besides, certain spatial transformations in the virtual try-on applications and image editing applications sometimes lead to chromatic loss in the output image. In such cases, color transfer methods are used to transfer the source color on the output image. 

References:

  1. Color Transfer between Images by Reinhard et al, 2001.
  2. Super-fast color transfer between images by Adrian Rosebrock ,2014
  3. Histogram specification by Ross Moore, Mathematics Department, Macquarie University, Sydney.

Histogram Specification: A Fast and Flexible Method to Process Digital Images