Two-Stage AI Image Processing for Real-Time Color Constancy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning-based neural networks face performance degradation and real-time processing limitations when predicting color constancy under extreme lighting changes, requiring significant computation resources.
Innovation Solution
An electronic device employs a first AI model trained using a second AI model, which is trained by generating predicted images under various lighting conditions and adjusting parameters based on loss values, to correct lighting conditions and maintain color constancy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep learning-based neural network is used to predict color constancy under arbitrary lighting conditions, then color restoration performance is improved, but computation resource requirements increase significantly and real-time processing is limited
Solution Approach 1:
The patent segments the color constancy prediction task into two distinct stages: a training stage where a second AI model learns color restoration from sample images under multiple lighting conditions, and an inference stage where a first AI model quickly predicts lighting conditions and restores colors in real-time. This segmentation allows the computationally intensive learning to occur offline while enabling real-time application online.
Solution Approach 2:
The patent applies preliminary action by pre-training the second AI model on diverse sample images collected under multiple lighting conditions before actual use. This preliminary training creates a knowledge base that the first AI model can leverage during real-time processing, eliminating the need for complex computations during inference and enabling real-time color constancy prediction.
2Measurement precision
If a deep learning-based neural network is used to predict color constancy under extreme lighting changes, then color restoration performance is improved, but prediction performance degrades under extreme conditions
Solution Approach 1:
The patent changes the parameter of training data diversity by collecting sample images under multiple varying lighting conditions during the training phase. This diverse training enables the second AI model to learn robust color restoration patterns that generalize to extreme lighting conditions, improving prediction reliability without sacrificing accuracy.
Solution Approach 2:
The patent incorporates feedback mechanisms during training by comparing predicted images with actual images under known lighting conditions to calculate loss values. This feedback loop allows the second AI model to iteratively adjust its parameters, optimizing its ability to handle extreme lighting changes and improving overall prediction reliability.
3Measurement precision
If significant computation resources are allocated to implement a high-performance model, then color restoration accuracy is improved, but real-time processing is limited
Solution Approach 1:
The patent segments computational work into two phases: an offline training phase where the second AI model is trained on diverse lighting conditions with high computational resources, and an online inference phase where the first AI model performs quick predictions using the pre-learned patterns. This segmentation transfers computational burden from runtime to training time, enabling real-time processing.
Solution Approach 2:
The patent performs preliminary training of the second AI model on diverse lighting conditions before actual application. This preliminary action creates optimized prediction models that can operate quickly during real-time use, significantly reducing processing time while maintaining high color restoration accuracy through the knowledge gained during training.
Data Source
AI summary
An electronic device includes one or more processors configured to collect an original image for a target scene, and generate a corrected modified image by reflecting a predicted lighting condition of the original image in response to inputting the original image to a first artificial intelligence (AI) model, wherein the first AI model is generated based on a second AI model trained based on any one or any combination of any two or more of a sample image collected under a plurality of lighting conditions, a first actual image comprising a lighting condition of the sample image, and a second actual image comprising color information of an object in the sample image.