Image Processing Apparatus Hardware Implementation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing super-resolution processes using learned databases face challenges with large circuit size, memory capacity, and noise immunity, making them unsuitable for hardware implementation and prone to picture quality deterioration due to noise.
Innovation Solution
An image processing apparatus that separates feature and non-feature components of low-resolution images, uses coefficient data to convert them to high-resolution, and averages pixel values across overlapping patches to generate a high-resolution output, reducing circuit size and memory while improving noise immunity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If similarity calculation is performed for each dimension of the search vector for each patch in the learned database, then high-resolution prediction accuracy is improved, but the amount of calculation per patch increases and circuit size becomes large
Solution Approach 1:
The patent segments the large learned database into multiple small sub-databases, each storing coefficient data for specific patch patterns. Instead of performing similarity calculations across the entire database, the system divides the search space into manageable segments, reducing computational complexity while maintaining prediction accuracy.
Solution Approach 2:
The patent applies local quality by creating pattern-specific coefficient data for different types of patches (e.g., edge regions, flat regions, texture regions). Each sub-database contains optimized coefficient data tailored to specific local characteristics, allowing accurate prediction without requiring exhaustive similarity calculations across all possible patterns.
2Measurement precision
If a large number of patches are stored in the learned database to improve prediction accuracy, then high-resolution quality is improved, but the required memory capacity becomes large
Solution Approach 1:
The patent segments the learned database into multiple small sub-databases organized by patch patterns. Instead of storing all possible high-resolution patches in one large database, the system creates specialized sub-databases for different pattern types, significantly reducing total memory requirements while maintaining comprehensive coverage of image variations.
Solution Approach 2:
The patent uses lightweight coefficient data representations instead of storing complete high-resolution patch images. The coefficient data serves as a compact mathematical model that can be applied to generate high-resolution patches on-demand, replacing the need to store large amounts of actual image data.
3Device complexity
If the number of dimensions of the search vector or the number of patches in the learned database is reduced to decrease circuit size or memory capacity, then hardware implementation becomes feasible, but noise immunity deteriorates
Solution Approach 1:
The patent enhances noise immunity through local quality by creating dedicated sub-databases for different patch patterns including noise-resistant representations. Each sub-database contains coefficient data optimized for specific local characteristics, allowing the system to maintain robust noise immunity even with reduced search vector dimensions by leveraging pattern-specific knowledge.
4Quantity of substance
If the number of patches in the learned database is reduced to decrease memory capacity, then hardware implementation becomes feasible, but the problem of noise-induced deterioration in picture quality becomes more prominent
Solution Approach 1:
The patent addresses this contradiction by organizing the reduced database into pattern-specific sub-databases. Each sub-database contains coefficient data optimized for specific local image characteristics, ensuring that even with fewer total patches, the system maintains high picture quality by selecting the most appropriate pattern-matched coefficients for each local region.
Data Source
AI summary
An image processing apparatus and image processing method enlarge an input image (Din) to generate a low-resolution enlarged image (D101). Depending on a result of identification of the pattern of the input image (Din), coefficient data (D108) for conversion to a high-resolution are selected, and a feature component (D102H) of a low resolution is converted to a feature component (D103H) of a high resolution. Decision as to whether or not the pattern of a local region of the input image (Din) is flat is made. If it is flat, the coefficient data are so selected that no substantial alteration is made to pixel values in a high-resolution conversion unit (103). It is possible to reduce the circuit size and the memory capacity, and to improve the noise immunity, and achieve conversion to a high resolution suitable for implementation by hardware.