Image Processing Apparatus Patch-Based Neural Network Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional noise reduction methods using neural networks are ineffective in reducing noise uniformly across various image contents, leading to residual noise in images with textures and complexity.
Innovation Solution
An image processing apparatus that segments input images into patches and selectively uses either an original image estimation neural network or a noise estimation neural network based on the content of each patch, integrating the processed patches to achieve robust noise reduction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If a conventional neural network method is used for noise reduction, then the processing speed is improved, but the noise reduction effect deteriorates in images with textures
Solution Approach 1:
The image is divided into multiple patches, and each patch is processed by selecting the most appropriate neural network (noise estimation NN or original image estimation NN) based on its local characteristics. This segmentation allows different processing strategies to be applied to different regions, resolving the contradiction between speed and noise reduction effect.
Solution Approach 2:
Different neural networks are selected and applied to different patches based on their local content characteristics. Patches with simple characteristics use one type of neural network, while patches with complex characteristics use another type, ensuring optimal noise reduction for each local region while maintaining overall processing efficiency.
2Manufacturing precision
If a dedicated neural network for specific texture is used, then the noise reduction effect is improved for that texture, but the device complexity and learning difficulty increase
Solution Approach 1:
Two types of neural networks are trained in advance and stored in the system. During processing, the system universally applies the appropriate network type based on patch characteristics without requiring complex dedicated networks for each specific texture. This multi-functionality approach reduces learning difficulty while maintaining effective noise reduction.
3Device complexity
If a single neural network is used for all images, then the device complexity is reduced, but the adaptability to different image contents deteriorates
Solution Approach 1:
The system dynamically selects the appropriate neural network type for each patch based on its characteristics. This dynamic adaptation allows the system to handle diverse image contents effectively without requiring a single overly complex network, balancing adaptability with manageable system complexity.
Data Source
AI summary
Noise is adequately reduced irrespective of the content of an input image. In an embodiment of the present invention, an image processing apparatus that executes noise reduction processing of an image includes: a first estimation unit that estimates noise contained in the image; a second estimation unit that estimates an original image, which is the image from which the noise is removed; a noise reduction unit that performs the noise reduction processing on each of partial areas of the image by using the first estimation unit or the second estimation unit depending on the contents of the partial areas; and an integration unit that integrates the partial areas on which the noise reduction processing is performed.