Image Generation Learning With Shared Noise to Prevent Blurring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image generation devices using machine learning to reduce noise in training input images result in image smoothing and blurring due to the learning of filters for low-frequency noise reduction.
Innovation Solution
An image generation device that adds the same noise to both training input and output images, learns a model to extract or remove specific portions through machine learning, and generates images using this model to suppress blurring.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a filter for extracting the low-frequency area of the image is learned to smooth noise in the training input image, then noise reduction is improved, but the output image becomes blurred
Solution Approach 1:
The patent extracts and removes the specific portion (e.g., bone, blood vessel, or other target structures) from the image rather than smoothing the entire image. By focusing the learning model on extracting specific structures rather than general noise reduction, the method avoids the blurring effect that occurs when low-frequency filters are applied to the whole image.
Solution Approach 2:
Instead of approaching noise reduction by smoothing the entire image (conventional approach), the patent inverts the approach by adding noise to both input and output images and learning to extract the difference. This inversion allows the model to focus on structural extraction without being biased toward smoothing operations.
2Measurement precision
If machine learning is performed to reduce noise in the training input image, then image quality is improved, but the learning model learns to smooth the image causing blurring
Solution Approach 1:
The patent converts the harmful effect of noise into a beneficial training mechanism. By adding noise to both the training input image and training output image, the method creates a learning scenario where the model must distinguish between the specific portion and noise. This transforms noise from a detrimental factor into a useful training element that prevents the model from learning smoothing filters.
Solution Approach 2:
The patent performs preliminary action by adding noise to both images before training begins. This preliminary noise addition ensures that the learning model is trained from the outset to handle noisy conditions and focus on structural extraction rather than smoothing, preventing blurring from the beginning of the training process.
3Manufacturing precision
If the same noise is added to both training input image and training output image, then the learning model can extract specific portions without smoothing, but the training process becomes more complex
Solution Approach 1:
The patent applies parameter changes by systematically varying the noise characteristics (e.g., noise type, noise level, noise distribution) during training. By adjusting these noise parameters, the method creates diverse training scenarios that teach the model to robustly extract specific portions under different noisy conditions, improving generalization without requiring overly complex training procedures.
Data Source
AI summary
This image generation device is provided with: a learning image generation unit (1) for generating a training input image and a training output image based on three-dimensional data; a noise addition unit (2) for adding the same noise to the training input image and the training output image; a learning unit (3) for learning a learning model for extracting or removing a specific portion by performing machine learning based on the training input image to which the noise has been added and the training output image to which the noise has been added; and an image generation unit (4) for generating an image from which the specific portion has been extracted or removed by using a learned learning model.