Differentiable Histogram Loss Functions for Intensity-based Image-to-Image Translation.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image-to-image translation is a challenging task in computer vision.
• Existing methods often struggle with precise control over output image appearance and color distribution.
• Generative neural networks require effective loss functions to guide the synthesis process.

Purpose of the Study:

• To introduce HueNet, a novel deep learning framework for differentiable histogram construction.
• To present HueNet's applicability to paired and unpaired image-to-image translation problems.
• To demonstrate HueNet's effectiveness in tasks requiring predefined output image colors.

Main Methods:

• Augmenting generative neural networks with histogram layers.
• Defining new histogram-based loss functions: Earth Mover's Distance for color similarity and mutual information for structural similarity.
• Applying the framework to color transfer, exemplar-based image colorization, and edges-to-photo tasks.

Main Results:

• HueNet enables differentiable construction of 1D and 2D histograms.
• The framework successfully constrains structural appearance and color distribution of synthesized images.
• Demonstrated strong performance on color transfer, exemplar-based colorization, and edges-to-photo tasks.

Conclusions:

• HueNet offers a novel approach to image-to-image translation using histogram-based constraints.
• The framework provides enhanced control over output image characteristics.
• HueNet advances generative modeling for specific image synthesis applications.