Wide Angle-of-View Image Generation Luminance Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for generating wide angle-of-view images using machine learning models struggle to accurately estimate high luminance portions, leading to unnatural placement or absence of these elements in the generated images, affecting the realism of light source settings in Image-Based Lighting (IBL).
Innovation Solution
A learning device and method that includes an input image acquisition section, a wide angle-of-view image generation section, and a learning section that updates parameter values based on comparisons between generated and comparative images, focusing on luminance levels to enhance the accuracy of high luminance portion estimation in wide angle-of-view images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a machine learning model is used to generate wide angle-of-view images from common camera images, then the need for specialized equipment and knowledge is eliminated, but the accuracy of high luminance portion estimation deteriorates
Solution Approach 1:
The patent changes the parameter weighting in the loss function by introducing luminance-level-dependent weights. Pixels with high luminance values are assigned larger weights in the loss calculation, forcing the model to pay more attention to accurately estimating high luminance portions during training. This resolves the contradiction by modifying the training parameters to prioritize accurate estimation of critical high luminance regions while still using common camera images for training.
2Device complexity
If uniform parameter updates are applied during learning, then the learning process is simple, but the estimation accuracy of high luminance portions deteriorates
Solution Approach 1:
The patent applies local quality by differentiating the update amount of parameter values based on pixel luminance levels. Instead of uniform updates, the system calculates different update amounts for different regions: pixels with high luminance receive larger update amounts while other pixels receive smaller updates. This localized approach improves high luminance estimation accuracy without significantly increasing overall system complexity.
3Reliability
If the model focuses on general image generation, then overall image quality is maintained, but high luminance portion placement becomes unnatural
Solution Approach 1:
The patent modifies the loss function parameters by introducing luminance-dependent weighting factors. The loss function is designed to calculate different error weights based on pixel luminance levels, ensuring that high luminance portions contribute more significantly to the gradient updates. This parameter change enables the model to prioritize accurate high luminance portion placement while maintaining overall image generation quality.
Data Source
AI summary
A second learning data acquisition section acquires an input image. A wide angle-of-view image generation section generates, in response to an input of the input image, a generated wide angle-of-view image that is an image having a wider angle of view than the input image. The second learning data acquisition section acquires a comparative wide angle-of-view image that is an image to be compared with the generated wide angle-of-view image. A second learning section performs learning for the wide angle-of-view image generation section by, on the basis of a comparison result between the generated wide angle-of-view image and the comparative wide angle-of-view image, updating parameter values of the wide angle-of-view image generation section such that, according to the luminance levels of pixels in the comparative wide angle-of-view image or the luminance levels of pixels in the generated wide angle-of-view image, update amounts of the parameter values concerning the pixels are increased.