Image Reconstruction Using Cross-Pooling for Long-Range Dependencies
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image reconstruction methods based on deep learning fail to capture long-distance dependencies between different image regions, limiting their effectiveness in capturing complex image structures.
Innovation Solution
The method employs cross-pooling techniques involving one-dimensional convolution and vector duplication to construct feature maps that capture long-distance dependencies, followed by two-dimensional convolution and feature mapping to generate a reconstructed image, reducing computational complexity while enhancing image reconstruction performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If conventional deep learning methods (unfolding or non-unfolding) are used for image reconstruction, then the reconstruction process can be completed, but the long-distance dependencies between different image regions cannot be captured
Solution Approach 1:
The patent segments the feature map construction into distinct directional components: horizontal feature maps capturing row-wise dependencies and vertical feature maps capturing column-wise dependencies. This segmentation allows the model to systematically capture long-distance dependencies in both dimensions separately, then fuse them to achieve comprehensive long-range feature capture without overwhelming complexity.
Solution Approach 2:
The patent transitions from standard 2D convolution to 1D convolution applied in specific directions (horizontal and vertical). By changing the convolution dimensionality and orientation, the model can effectively capture long-distance dependencies along rows and columns without the computational burden of full 2D long-range operations, resolving the contradiction between information capture and complexity.
2Reliability
If more feature maps are generated to improve network performance, then the reconstruction quality improves, but the computational complexity increases
Solution Approach 1:
The patent divides the feature map generation process into segmented directional operations (horizontal and vertical 1D convolutions). This segmentation enables the generation of multiple feature maps with different directional sensitivities, improving reconstruction quality through diverse feature representation while maintaining computational efficiency by avoiding full 2D convolution operations.
Solution Approach 2:
Instead of applying full 2D convolution operations which would be computationally expensive, the patent applies 1D convolution operations partially in horizontal and vertical directions. This partial action approach generates sufficient feature diversity for high-quality reconstruction while significantly reducing computational burden compared to exhaustive 2D operations.
Data Source
AI summary
The present application relates to image processing technical field, and provides a method for image reconstruction, an apparatus, a terminal device, and a storage medium. The method first extracts an initial feature map of an original image, then calculates an average value of each column pixel in the initial feature map, and constructs a target row vector and duplicates the target row vector in the column direction after convolution processing, to obtain a feature map. In addition, an average value of the element of each row of pixels in the initial feature map is calculated respectively, and a target column vector is constructed. It is duplicated in a row direction to obtain another feature map, and then the two feature maps are fused. Finally, two-dimensional convolution processing is performed on a fused feature map, and a reconstructed image is generated, thereby the long-distance dependencies of the image can be captured.