Surround View Image Stitching Feedback for Seam Quality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image stitching technologies struggle to create seamless and realistic composite representations from multiple camera views in complex environments, often resulting in visible seams, misalignment, and artifacts that do not align with viewer subjective quality.
Innovation Solution
A stitching quality assessment metric is used to determine objective quality measures, which are then fed into a machine learning classifier to infer subjective quality, allowing adjustments to stitching parameters for improved visual quality, using techniques like blending and alpha maps to minimize seams and artifacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional image stitching algorithms are used to combine multiple camera views, then the stitching process can be completed, but visible seams and image artifacts appear in the composite representation
Solution Approach 1:
The patent implements a feedback mechanism where a machine learning classifier evaluates the stitching quality of composite images and provides feedback to adjust stitching parameters. The classifier analyzes objective quality measures (such as seam visibility and artifact presence) and uses this information to iteratively optimize the stitching process, thereby reducing seams and artifacts while improving overall stitching quality
Solution Approach 2:
The patent employs parameter changes by adjusting stitching parameters (such as blending weights, transformation matrices, and overlap regions) based on feedback from the machine learning classifier. The system dynamically modifies these parameters to minimize the presence of seams and artifacts in the composite image, transforming the stitching process from a static algorithm to an adaptive optimization system
2Reliability
If stitching parameters are adjusted to reduce visible seams, then seam visibility improves, but processing time and computational complexity increase
Solution Approach 1:
The patent applies preliminary action by pre-training the machine learning classifier on a dataset of stitched images with various quality levels. This pre-training enables the classifier to quickly evaluate stitching quality and provide feedback without requiring extensive computation during the actual stitching process, thus reducing processing time while maintaining the ability to optimize seam visibility
Solution Approach 2:
The patent implements partial action by focusing the optimization process on specific critical regions of the composite image where seams are most likely to appear (such as overlap regions between adjacent camera views). Rather than uniformly processing the entire image, the system concentrates computational resources on these critical areas, achieving good seam reduction with reduced overall processing time
3Area of stationary object
If multiple camera views are combined to create a comprehensive environment representation, then the field of view increases, but alignment accuracy decreases due to complex scene variations
Solution Approach 1:
The patent applies dynamics by using dynamic feature matching that adapts to different scene complexities. The machine learning classifier learns to identify and match relevant features across multiple camera views, adjusting its matching criteria based on the specific characteristics of the scene (such as texture, lighting conditions, and geometric structures). This dynamic approach maintains alignment accuracy even as the field of view expands to include more diverse and complex environments
Data Source
AI summary
Stitching of multiple images into a composite representation can be performed using a set of stitching parameters determined based, at least in part, upon a subjective stitching quality assessment value. A stitched image can be compared against its constituent images to obtain one or more objective quality metrics. These objective quality metrics can be fed, as input, to a trained classifier, which can infer a subjective quality assessment metric for the stitched (or otherwise composited) image. This subjective quality assessment metric can be used to adjust one or more compositing parameter values in order to provide at least a minimum subjective quality assessment value for composited images.