Real-Time Deep Image Retouching Based on Learnt Semantics Dependent Global Transforms.

Area of Science:

• Computer Vision
• Machine Learning
• Digital Image Processing

Background:

• Photo retouching by artists is complex and difficult to model analytically.
• Existing methods for image enhancement often involve end-to-end pixel mapping, which can be computationally intensive.

Purpose of the Study:

• To develop a computationally efficient method for mimicking artists' photo retouching techniques.
• To simplify the machine learning problem of image transformation using a novel deep learning architecture.

Main Methods:

• Modeling artistic retouching effects using parametric tone mapping and affine chrominance transforms weighted by saliency.
• Constructing a Deep Artful Image Transform (DAIT) using convolutional neural networks (CNN) to learn image-dependent transform parameters.
• Reducing computational complexity by learning transform parameters instead of end-to-end pixel mapping.

Main Results:

• The DAIT approach significantly reduces neural network computation complexity by two orders of magnitude.
• Improved robustness and generalization capabilities during the inference stage.
• Demonstrated suitability for real-time video enhancement with simple temporal processing.

Conclusions:

• The proposed DAIT method effectively models and mimics artistic photo retouching with high efficiency.
• The approach offers a robust and generalizable solution for image and video enhancement.
• DAIT provides a significant advancement in computational photography and machine learning applications.