Neural Network Image Blending via Confidence Value Thresholding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Image artifacts occur when blending two or more images due to differences in resolution and display configurations, leading to suboptimal results and the need for cumbersome adjustments to remove artifacts, especially in video conferencing applications.
Innovation Solution
A system that uses one or more neural networks to generate confidence values for each pixel in an image, allowing for the blending of images while avoiding confidence values that would cause artifacts by training the neural network to predict extremely high confidence values and penalizing predictions below a certain threshold.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If two or more images are blended together based on confidence values, then image blending is achieved, but image artifacts occur due to different resolution configurations
Solution Approach 1:
The patent changes the parameter range of confidence values by training the neural network to output values outside the artifact-causing range (e.g., below a first threshold or above a second threshold). This parameter transformation resolves the contradiction by maintaining blending capability while eliminating artifacts through controlled value distribution.
Solution Approach 2:
The patent implements feedback mechanisms during neural network training by penalizing predictions that fall within the artifact-causing confidence value range. The loss function provides feedback to adjust network parameters, ensuring future predictions avoid the harmful range while maintaining effective blending performance.
2Object-affected harmful factors
If confidence values are adjusted to remove artifacts, then image quality improves, but processing time increases due to latency
Solution Approach 1:
The patent performs preliminary action by training the neural network in advance to predict confidence values that are already outside the artifact-causing range. This pre-training ensures that during actual image blending operations, no additional processing or adjustment is needed, thereby eliminating latency while maintaining artifact-free output.
Solution Approach 2:
The patent transforms the confidence value parameter space during training so that the network naturally outputs values in ranges that avoid artifacts. This parameter change is baked into the model during training, allowing rapid inference without post-processing adjustments, thus resolving the time-quality tradeoff.
3Object-affected harmful factors
If neural network predictions are penalized below a threshold, then confidence values are optimized to avoid artifacts, but training complexity increases
Solution Approach 1:
The patent modifies the training process by changing the loss function parameters to penalize predictions within the artifact-causing range. This parameter adjustment in the objective function guides the network to learn appropriate confidence value distributions without requiring complex architectural modifications or multi-stage training procedures.
Data Source
AI summary
Apparatuses, systems, and techniques to blend two or more images based on confidence values of objects within said two or more images. In at least one embodiment, one or more confidence values in one or more images are generated using one or more neural networks that are used, for example, to blend two or more images to be displayed.