Image Processing Apparatus Parameter Segmentation for Correction Conditions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing techniques struggle to maintain inference accuracy when image characteristics change due to different correction conditions during learning and inference, leading to improper noise removal and other image processing issues.
Innovation Solution
An image processing apparatus that performs machine learning by processing training and teacher images under various correction conditions, storing parameters associated with these conditions, and selecting the appropriate parameters for inference based on the current correction conditions to ensure accurate image processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If one kind of network parameters is used to perform inference for correcting blur caused due to aberration or diffraction, then the device complexity is reduced, but the measurement precision of image processing deteriorates when correction conditions differ between learning and inference
Solution Approach 1:
The patent segments the network parameters into multiple sets, each corresponding to a specific correction condition (e.g., different aberration correction settings). Instead of using a single unified parameter set, the system divides parameters into condition-specific groups, allowing accurate inference for each correction scenario while managing complexity through organized segmentation.
Solution Approach 2:
The patent implements dynamic parameter selection where the network parameters are changed based on the actual correction condition being applied. The system dynamically switches between different parameter sets depending on the correction setting, making the inference process adaptive to varying conditions rather than static.
2Adaptability or versatility
If image processing changes characteristics of an image depending on correction condition, then the adaptability of the system is improved, but the reliability of inference deteriorates when correction conditions mismatch between learning and inference
Solution Approach 1:
The patent performs preliminary action by pre-training the neural network for each specific correction condition before actual inference. Multiple network parameters are trained in advance under different correction settings (e.g., different aberration corrections), so when inference is performed, the system can select the pre-trained parameters that match the current correction condition, ensuring reliability.
Solution Approach 2:
The patent changes parameters based on correction conditions by selecting different network parameter sets corresponding to different correction settings. When the correction condition changes (e.g., switching aberration correction levels), the system changes the network parameters to match, maintaining consistency between learning and inference conditions.
3Measurement precision
If multiple parameters are stored for different correction conditions, then the measurement precision of inference is improved, but the device complexity increases due to parameter management
Solution Approach 1:
The patent adds another dimension to parameter storage by organizing parameters not just as a flat set but with an additional correction condition dimension. Parameters are stored in a structured manner where each parameter set is tagged with its corresponding correction condition, allowing efficient retrieval and management without linearly increasing complexity.
Data Source
AI summary
An information processing apparatus that performs machine learning of a learning model for improving the accuracy of inference of an image even in a case where the characteristics of an image are changed by correction applied to the image. The image processing apparatus includes an image processor configured to perform image processing on a training image and a teacher image according to each of a plurality of correction conditions. Machine learning of the learning model is performed using the training image and the teacher image. A plurality of parameters obtained by performing the machine learning of the learning model are stored in association with the plurality of correction conditions, respectively.