Heterogeneous Image Pose Registration Using U-Net Phase Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for pose estimation and registration of heterogeneous images are not ideal due to differences in angle, proportion, and viewing angle, and are hindered by illumination and occlusion, making accurate matching difficult.
Innovation Solution
A neural network-based method using pre-trained U-Net networks and phase correlation algorithms to extract isomorphic features and perform Fourier and logarithmic polar coordinate transforms to estimate rotation, scaling, and translation transforms between heterogeneous images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional pose estimation methods are used for heterogeneous images, then the processing speed is fast, but the matching accuracy is poor due to differences in angle, proportion and viewing angle
Solution Approach 1:
The patent applies parameter changes by transforming images from Cartesian coordinates to logarithmic polar coordinates, which converts rotation and scaling operations into translation operations. This coordinate transformation enables accurate pose estimation for heterogeneous images by normalizing the geometric differences between images, thereby improving matching accuracy without significantly increasing methodological complexity
Solution Approach 2:
The patent introduces an intermediary approach by using a pre-trained U-Net network as a feature extractor that bridges the gap between heterogeneous images. The network extracts isomorphic features from images with different angles, proportions and viewing angles, enabling accurate matching while maintaining computational efficiency
2Reliability
If traditional feature matching methods are used, then the computational time is short, but the reliability is low due to illumination and occlusion disturbances
Solution Approach 1:
The patent implements feedback mechanisms through the pre-trained U-Net network that continuously refines feature extraction based on the input images. The network's pre-training on diverse image datasets enables it to robustly handle illumination changes and occlusions, improving reliability while maintaining fast processing through efficient feature representation
Solution Approach 2:
The patent applies preliminary action by pre-training the U-Net network on a comprehensive dataset before actual pose estimation tasks. This pre-training equips the network with the ability to recognize and ignore disturbances like illumination changes and occlusions, ensuring reliable pose estimation in real-time without requiring complex real-time adjustments
3Measurement precision
If heterogeneous images are processed directly, then the workflow is simple, but the measurement precision deteriorates due to differences in angle, proportion and viewing angle
Solution Approach 1:
The patent uses parameter changes by applying logarithmic polar coordinate transformation to convert rotation and scaling differences into translation problems. This mathematical transformation simplifies the matching process while significantly improving measurement precision for heterogeneous images with different geometric parameters
Solution Approach 2:
The patent introduces an intermediary feature extraction stage using the pre-trained U-Net network that processes heterogeneous images before matching. This intermediary layer normalizes the geometric differences and extracts isomorphic features, enabling accurate pose estimation while maintaining operational simplicity through automated processing
Data Source
AI summary
The present invention relates to the field of image processing. Disclosed are a neural network-based pose estimation and registration method for heterogeneous images. In the present invention, a phase correlation algorithm is optimized to be differentiable and embedded into an end-to-end learning network framework, and a neural network-based pose estimation method for heterogeneous images is constructed. According to the method, an optimal feature extractor can be found for a result of image matching, a solution can be obtained without exhaustive evaluation, and good interpretability and generalization capability are achieved. The test results show that the present invention allows for accurate pose estimation and registration for heterogeneous images and shortening of the required time, has high accuracy and real-time performance, can meet actual application requirements, and can be applied in fields such as robot self-positioning.