Neural Image Color Transformation Without Reference Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network models face performance degradation due to domain differences caused by seasonal variations or camera characteristics, necessitating color transformation to address this issue, and existing methods like 'Color Transfer Using a Probabilistic Moving Least Squares' incur high computational complexity.
Innovation Solution
A neural network model is configured with an operation block for white balancing, a correction block for color correction, and a mapping block for applying a lookup table, utilizing depth-wise convolution and batch normalization layers, along with a comparison model for unsupervised learning to transform images without a reference image, reducing computational complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If color transformation is performed using existing methods like 'Color Transfer Using a Probabilistic Moving Least Squares', then color transformation can be achieved, but computational complexity increases
Solution Approach 1:
The patent replaces traditional mechanical/mathematical color transformation methods (probabilistic moving least squares) with a neural network-based system. The neural network model learns color transformation patterns automatically from training data, substituting complex iterative mathematical optimization with a pre-trained computational model that performs transformation through forward propagation, significantly reducing real-time computational complexity.
Solution Approach 2:
The patent performs color transformation computations in advance during the training phase. The neural network is trained on paired training images to learn the mapping between source and target color spaces beforehand. This preliminary learning allows the model to perform rapid transformations during deployment without requiring complex real-time calculations, effectively moving the computational burden from runtime to training time.
2Reliability
If training data in a plurality of domains is collected to train neural network models, then model performance can be improved, but it becomes difficult to directly collect training data due to domain differences
Solution Approach 1:
The patent introduces a color transformation system as an intermediary to bridge domain differences in training data. By transforming images from different domains (different seasons, cameras, lighting conditions) into a unified color space, the system creates compatible training data without requiring direct collection from multiple domains. This intermediary transformation enables the neural network to learn from diverse data sources while maintaining consistency.
Solution Approach 2:
The patent changes the color parameters of images through automated transformation to adapt training data across different domains. By adjusting color space parameters (converting between RGB, LAB, YCrCb spaces and applying learned transformation matrices), the system standardizes images from various sources, making them suitable for unified training without manual intervention or direct multi-domain data collection.
3Reliability
If a reference image is needed for color transformation, then color transformation can be performed, but optimization issues arise and computational complexity increases
Solution Approach 1:
The patent enables the color transformation system to serve itself by automatically learning transformation parameters from training data without requiring manual reference images or iterative optimization. The neural network automatically adapts to different domains and learns the optimal transformation mappings during training, eliminating the need for external reference images and complex real-time optimization processes.
Data Source
AI summary
A method with image transformation includes: identifying an original image; and determining a transformed image by inputting the original image to a neural network model configured to transform a color of the original image, wherein the neural network model comprises an operation block configured to perform white balancing on the original image, a correction block configured to correct a color of an output image of the operation block, and a mapping block configured to apply a lookup table to an output image of the correction block.