Generative Adversarial Networks for Local Noise Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques face difficulties in effectively removing local noise from images, particularly when training machine learning models to minimize errors across entire images with noise regions.
Innovation Solution
The method involves estimating noise regions in images using machine learning converters and weighting these regions to generate noise-free images for training, allowing for effective noise removal through a process involving neural network models and generative adversarial networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional machine learning training minimizes errors across entire images, then overall image quality is improved, but local noise removal effectiveness deteriorates
Solution Approach 1:
The patent segments the image processing task by introducing a mask that divides the image into noise regions and non-noise regions. The loss function is then applied separately to different regions, allowing localized optimization for noise removal while maintaining overall image quality. This segmentation enables the model to focus computational effort on specific problem areas rather than treating the entire image uniformly.
Solution Approach 2:
The patent implements local quality by applying different loss weights to different regions of the image. Specifically, the loss function assigns higher weights to noise regions identified by the mask, ensuring that the model prioritizes accurate noise removal in these areas. This local quality approach allows the system to optimize for noise removal effectiveness in problematic regions while maintaining acceptable overall image quality.
2Quantity of substance
If machine learning models are trained on images with noise regions, then training data availability is improved, but training accuracy deteriorates due to error minimization across entire images
Solution Approach 1:
The patent extracts the noise region information from the training images by using a pre-trained noise detection model to generate masks. These masks are then used to create weighted loss functions that focus training attention on noise regions. This extraction allows the system to utilize noisy training data effectively by isolating and prioritizing the problematic regions during the training process, thereby maintaining both data availability and training accuracy.
3Ease of manufacture
If uniform loss weighting is applied across entire images, then training simplicity is improved, but local noise removal performance deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-processing the training images to generate noise region masks before the main training process. These masks are created using a pre-trained noise detection model, which identifies noise regions in advance. The pre-generated masks are then integrated into the loss function to guide the training process, allowing the model to focus on noise regions without adding significant complexity to the training procedure.
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
An information processing method includes: obtaining noise region estimation information output from a first converter (30) by a first image including a noise region being input to the first converter (30); obtaining a second image, on which noise region removal processing has been performed, output from a second converter (60) by the noise region estimation information and the first image being input to the second converter (60); generating a fourth image including the estimated noise region by using the noise region estimation information and a third image including no noise region and a scene corresponding to the first image; training the first converter (30) by using machine learning in which the first image is reference data and the fourth image is conversion data; and training the second converter (60) by using machine learning in which the third image is reference data and the second image is conversion data.