Image Registration via Joint Spatial Gradient Maximization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image registration methods fail to effectively register multi-modal images, such as visible and infrared images, due to their different characteristics and complex relationships between pixel intensities, especially when large motion and viewpoint changes occur, as they assume small displacements and high feature correlation.
Innovation Solution
A method that uses a joint gradient similarity function applied to pixels with large spatial gradient magnitudes, which are indicative of depth discontinuities, to maximize a similarity function through a gradient ascent procedure, allowing for registration without explicit initialization and handling large displacements, and is applicable to affine-based 2D image registration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional intensity-based feature extraction is used, then feature matching can be performed, but it fails for multi-modal images where features appear in one image but not in others (e.g., IR images of paintings appear homogenous)
Solution Approach 1:
The patent changes the parameter used for feature extraction from intensity values to spatial gradient magnitudes. By using gradient magnitude instead of intensity, the method captures geometric properties (edges, contours) that are invariant to intensity differences between modalities, enabling reliable feature extraction in multi-modal images where intensity-based methods fail.
2Adaptability or versatility
If mutual information is used for registering multi-modal images, then registration can handle different imaging modalities, but convergence is slow and requires complex optimization
Solution Approach 1:
The patent replaces the complex optimization process of mutual information maximization with a simpler gradient ascent procedure. By formulating the registration problem as maximizing the correlation between spatial gradient magnitudes, the method uses straightforward gradient-based optimization instead of complex iterative optimization, achieving fast convergence while maintaining adaptability to multi-modal images.
3Adaptability or versatility
If feature-based registration is used, then large motion and viewpoint changes can be handled, but extracting distinctive features invariant to illumination, scale and rotation is difficult
Solution Approach 1:
The patent changes the feature extraction parameter from intensity-based descriptors to spatial gradient magnitude. Gradient magnitude is inherently invariant to illumination changes (unlike intensity), and by using gradient-based features, the method naturally handles scale and rotation invariance, making feature extraction straightforward while maintaining capability for large motion and viewpoint changes.
4Manufacturing precision
If conventional direct methods are used, then pixel-to-pixel matching can be performed, but they require initialization and struggle with large displacements
Solution Approach 1:
The patent inverts the conventional approach by maximizing the correlation between spatial gradient magnitudes rather than minimizing intensity differences. This inversion allows the method to handle large displacements without requiring initialization, as the gradient-based similarity measure provides a robust starting point that naturally guides the registration process.
Data Source
AI summary
A method registers images by first determining spatial gradient magnitudes for each pixel in each image to provide a corresponding energy image. Each energy image is transformed according to motion parameters associated with the energy image. An average sum of the transformed energy images is maximized. The motion parameters are then updated according to the maximized average sum. The above steps are repeated until a termination condition is reached, and then the images can be registered with each other according to the updated motion parameters.