Image Denoising via Graph Penalty Minimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image denoising techniques often blur sharp edges and high-contrast areas while attempting to remove noise, reducing the diagnostic value of medical images and affecting the quality of photographic images.
Innovation Solution
A method that formulates a graph with nodes representing pixels and edge weights based on a penalty function that penalizes intensity changes between neighboring pixels, minimizing the total penalty to preserve image sharpness and optimize object curvature, using a penalty function defined as wij|xi−xj|+wik|xi−xk|−wjk|xj−xk|, where wij, wik, and wjk are curvature weights, and xi, xj, xk are pixel values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If simple averaging denoising is applied, then noise is removed, but blurring occurs at points of abrupt intensity change
Solution Approach 1:
The patent applies different processing treatments to different regions of the image based on local characteristics. The energy function evaluates local intensity variations and applies denoising selectively, preserving sharp edges where intensity changes abruptly while removing noise in uniform regions. This local differentiation resolves the contradiction by adapting the denoising strength to local image content.
Solution Approach 2:
The patent modifies the energy function parameters to balance noise removal and edge preservation. By adjusting the weighting between the data fidelity term and regularization term in the energy function, the system dynamically controls the trade-off between removing noise and maintaining sharpness, resolving the contradiction through parameter optimization.
2Object-affected harmful factors
If non-local means denoising is applied, then noise is removed by averaging similar areas, but computational complexity increases
Solution Approach 1:
The patent extracts and minimizes a specifically designed energy function that captures the essential denoising objective without requiring complex non-local comparisons. By formulating a tailored energy function with data fidelity and regularization terms, the system achieves effective noise removal while avoiding the computational burden of searching for similar patches across the entire image.
Solution Approach 2:
The patent simplifies the computational complexity by changing the approach from non-local patch matching to local energy minimization with adjustable parameters. The energy function parameters control the denoising behavior, replacing computationally intensive non-local operations with more efficient local computations that achieve similar results.
3Object-affected harmful factors
If TV denoising is applied, then noise is removed by minimizing total variation, but edges may be blurred or artifacts introduced
Solution Approach 1:
The patent improves edge accuracy by applying local quality principles through a modified energy function that evaluates local intensity relationships more precisely. The penalty function considers the relative ordering of pixel intensities in local neighborhoods, preserving edge structures more accurately than standard TV denoising while still removing noise effectively.
Solution Approach 2:
The patent inverts the conventional TV denoising approach by using a penalty function that rewards preservation of intensity ordering relationships rather than simply minimizing absolute variations. This inversion of the optimization objective helps preserve edges and geometric patterns more accurately while removing noise.
Data Source
AI summary
A method for removing noise from an image includes receiving image data including a plurality of pixels. A graph including a plurality of nodes and a plurality of edges interconnecting the nodes is formulated. Each pixel of the image data is represented as a node of the graph and each edge of the graph is assigned a weight based on a penalty function applied to the nodes connected by the edge where the penalty function is less when a value of a given pixel of the plurality of pixels is between or equal to the values of two neighboring pixels than when the value of the given pixel is either greater than or less than the values of both of the two neighboring pixels. A total penalty of the graph is minimized. A denoised image is provided based on the total penalty-minimized graph.