Sparsity Enforcing Neural Network for Image Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Image reconstruction is an ill-posed inverse problem due to limited hardware measurements, leading to underdetermined scenarios where existing methods like iterative reconstruction are computationally expensive and challenging to deploy effectively.
Innovation Solution
A Sparsity Enforcing Neural Network (SENN) is developed, utilizing a dual-projection function to enforce sparsity constraints on reconstructed images, which learns to reconstruct images from measurements by incorporating the observation matrix and employing sparsity-enforcing error-back propagation for efficient training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If iterative reconstruction is used to improve image reconstruction quality, then manufacturing precision is improved, but use of energy increases and productivity decreases
Solution Approach 1:
The patent replaces traditional iterative mechanical reconstruction algorithms with a neural network-based system. The neural network is trained offline to learn the inverse mapping from measurements to images, and during operation, it directly reconstructs images from measurements in a single forward pass, eliminating the need for iterative computations while maintaining high reconstruction quality.
Solution Approach 2:
The patent performs the computationally intensive work in advance by training the neural network offline using paired measurement-image data. Once trained, the network contains pre-learned reconstruction knowledge that can be rapidly applied to new measurements without requiring iterative computation at runtime, thus improving productivity while maintaining precision.
2Manufacturing precision
If iterative reconstruction is used to improve image reconstruction quality, then manufacturing precision is improved, but loss of time increases
Solution Approach 1:
The patent substitutes iterative computational algorithms with a trained neural network that performs reconstruction in a single forward pass. This eliminates the time-consuming iterative loops while preserving reconstruction quality, directly reducing computation time and time loss.
Solution Approach 2:
The neural network is trained beforehand on comprehensive datasets to learn optimal reconstruction patterns. During actual operation, the pre-trained network rapidly reconstructs images without requiring iterative refinement, significantly reducing the time loss associated with real-time computation.
3Manufacturing precision
If sparsity constraints are enforced using traditional methods to improve manufacturing precision, then manufacturing precision is improved, but device complexity increases
Solution Approach 1:
The patent replaces complex algorithmic sparsity enforcement mechanisms with a neural network that has been trained to implicitly learn and enforce sparsity constraints. The network's internal representations and learned features naturally capture sparse structures without requiring explicit constraint programming or complex optimization routines, thus reducing algorithmic complexity while maintaining reconstruction accuracy.
Data Source
AI summary
Systems and methods for a computer implemented image reconstruction system that includes an input interface to receive measurements of a scene. A memory to store a sparsity enforcing neural network (SENN) formed by layers of nodes propagating messages through the layers. Wherein at least one node of the SENN modifies an incoming message with a non-linear function to produce an outgoing message and propagates the outgoing message to another node of the SENN. Wherein the non-linear function is a dual-projection function that limits the incoming message if the incoming message exceeds a threshold. Such that, the SENN is trained to reconstruct an image of the scene from the measurements of the scene. A processor to process the measurements with the SENN to reconstruct the image of the scene. Finally, an output interface to render the reconstructed image of the scene.