Image Fusion Using Local and Global Weight Matrices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image fusion methods face challenges in efficiently combining images taken under different exposure settings, often resulting in high computational costs and artifacts due to non-linear transformations or operations limited to small local regions.
Innovation Solution
A method and system that determine local and global weight matrices for each source image, using feature matrices and similarity pyramids to combine images effectively, allowing for the creation of a fused image with improved dynamic range and reduced computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multi-scale decomposition methods (pyramid transform, wavelet transform) are used for image fusion, then image details and dynamic range are improved, but spatial computational cost increases
Solution Approach 1:
The patent divides the image fusion process into two independent stages: feature extraction stage (computing local feature matrices) and weight computation stage (computing final weight matrices). This segmentation allows parallel processing and reduces the computational burden on the spatial domain while preserving image details through multi-scale feature analysis.
Solution Approach 2:
The patent transforms the fusion problem from purely spatial domain operations to a combination of feature space and weight space operations. By computing local feature matrices and final weight matrices in a transformed domain, the method reduces spatial computational complexity while maintaining fusion quality through the relationship between features and weights.
2Measurement precision
If optimization-based methods computing local features and solving optimization problems are used, then fusion quality is improved, but temporal computational cost increases
Solution Approach 1:
The patent pre-computes local feature matrices for each input image before the fusion process. These feature matrices capture essential image characteristics and are stored for later use. During fusion, the method only needs to compute final weight matrices by combining pre-computed features, significantly reducing temporal computational cost while maintaining high fusion quality.
Solution Approach 2:
The patent creates local feature matrices as intermediate representations that copy essential image information into a compressed format. These feature matrices serve as templates that guide the fusion process, allowing the system to reuse computational results across different fusion operations and reduce real-time computational requirements.
3Measurement precision
If block-based optimization methods are used for image fusion, then local region fusion is improved, but spatial computational cost increases due to multiple small region operations
Solution Approach 1:
The patent merges the computations for multiple local regions into a unified process. Instead of separately optimizing each small block, the method computes local feature matrices across the entire image and then determines final weight matrices that coordinate all regions simultaneously. This merging approach reduces the total number of computational operations while maintaining local fusion quality through the structured relationship between features and weights.
Data Source
AI summary
A method and system is provided for combining information from a plurality of source images to form a fused image. The fused image is generated by the combination of the source images based on both local features and global features computed from the source images. Local features are computed for local regions in each source image. For each source image, the computed local features are further processed to form a local weight matrix. Global features are computed for the source images. For each source image, the computed global features are further processed to form a global weight vector. For each source image, its corresponding local weight matrix and its corresponding global weight vector are combined to form a final weight matrix. The source images are then weighted by the final weight matrices to generate the fused image.