Unsupervised CNN for Noisy Contour Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computer vision systems struggle with aligning noisy and partially occluded contours due to their sensitivity to noise and computational inefficiencies, especially in non-rigid registration scenarios where ground-truth warp fields are challenging to collect.
Innovation Solution
A deep convolutional neural network (CNN) is employed for unsupervised learning to progressively align noisy contours using a multiscale approach with progressively increasing complex transformations, coupled with a local shape-dependent Chamfer upper bound loss function that measures proximity and local shape similarity, making the system robust to noise and occlusions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If nonrigid registration approaches utilizing intensity based similarity metrics are used, then alignment flexibility is improved, but computational cost increases and sensitivity to corrupted pixels worsens
Solution Approach 1:
The patent segments the alignment process into multiple scales (coarse to fine) and divides the computational work across different resolution levels. By processing at downsampled scales first and progressively refining, the method reduces the total computational burden while maintaining alignment flexibility through the multiscale approach.
Solution Approach 2:
The patent performs preliminary alignment at coarse scales before refining at finer scales. This preliminary action at lower resolutions reduces the complexity of subsequent fine-grained alignment operations, thereby reducing overall computational cost while preserving alignment flexibility.
2Measurement precision
If deep learning based approaches with direct prediction of alignment warp at original scale are used, then alignment accuracy is improved, but robustness to noise and partial occlusions worsens
Solution Approach 1:
The patent segments the alignment process into multiple scales, starting with coarse-scale alignment and progressively refining to fine-scale alignment. This segmentation allows the network to learn alignment patterns at different levels of detail, improving robustness to noise at coarse scales while achieving high accuracy at fine scales.
Solution Approach 2:
The patent performs preliminary alignment at downsampled scales before final refinement at original scale. This preliminary action provides a robust initial alignment that is less sensitive to noise and occlusions, which then serves as a better starting point for the final high-accuracy alignment step.
3Ease of manufacture
If pixel level similarity metrics such as normalized cross correlation or mean squared error are used, then training simplicity is improved, but performance in noisy and partially observed contexts worsens
Solution Approach 1:
The patent moves from pixel-level comparison to contour-level comparison by extracting and matching contour representations. This dimensional change from 2D pixel grids to 1D contour curves makes the similarity metric more robust to noise and occlusions while maintaining training simplicity through the use of standard regression loss functions on the contour parameters.
Data Source
AI summary
Computer vision systems and methods for noisy contour alignment are provided. The system generates a loss function and trains a convolutional neural network with the loss function and a plurality of images of a dataset to learn to align contours with progressively increasing complex forward and backward transforms over increasing scales. The system can align an attribute of an image of the dataset by the trained neural network.