Image Property Map Filtering with Confidence-Weighted Neighborhood
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods for depth maps are prone to noise and errors, leading to degraded three-dimensional image quality, especially at transitions between image objects, due to noisy disparity estimation and lack of optimal performance in filtering techniques like bilateral filtering.
Innovation Solution
An apparatus and method that filters image property maps using a combined neighborhood image property value determined by a weighted combination of neighborhood image property values, where weights are dependent on confidence values and light intensity differences, allowing for asymmetric weighting based on confidence and intensity variations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If disparity estimation is used to generate depth values from stereo images, then depth information can be obtained, but the depth map becomes noisy and error-prone
Solution Approach 1:
The patent combines multiple depth maps from different view directions and merges them using a confidence-weighted approach. By integrating information from multiple stereo pairs and view directions, the system produces a more reliable final depth map that reduces noise and errors inherent in individual disparity estimations.
Solution Approach 2:
The system employs confidence maps that provide feedback about the reliability of each depth value. This confidence information is used to weight and select depth values during merging, allowing the system to rely more on accurate measurements and less on noisy estimates, thereby improving overall depth map quality.
2Reliability
If bilateral filtering is applied to improve depth map quality, then consistency and temporal stability improve, but the filtering does not provide optimum performance and artefacts remain
Solution Approach 1:
The patent applies different filtering strategies to different regions of the depth map based on local characteristics. By analyzing confidence values and depth discontinuities locally, the system adapts the filtering strength and method for each region, preserving edges where needed while smoothing noisy areas, thus achieving both consistency and accuracy.
Solution Approach 2:
The filtering approach is made dynamic by adjusting filter parameters based on local confidence values and depth variations. The system dynamically selects between different filtering modes and adjusts kernel sizes based on the local reliability of depth measurements, allowing optimal performance across diverse scene regions.
3Reliability
If confidence-weighted merging of multiple depth maps is performed, then depth map reliability improves, but computational complexity increases
Solution Approach 1:
The system performs partial merging by selectively combining depth maps based on confidence thresholds and region importance. Rather than merging all available depth maps uniformly, the system focuses computational resources on critical regions with lower confidence values, achieving improved reliability without proportionally increasing complexity.
Solution Approach 2:
The patent dynamically adjusts merging parameters such as confidence thresholds, kernel sizes, and weighting factors based on local depth map characteristics. By changing these parameters adaptively rather than using fixed values, the system achieves high reliability while managing computational complexity through efficient parameter selection.
Data Source
Figure 1
Figure 2
AI summary
An apparatus comprises receiver (101) receiving a light intensity image, confidence map, and image property map. A filter unit (103) is arranged to filter the image property map in response to the light intensity image and the confidence map. Specifically, for a first position, the filter unit (103) determines a combined neighborhood image property value in response to a weighted combination of neighborhood image property values in a neighborhood around the first position, the weight for a first neighborhood image property value at a second position being dependent on a confidence value for the first neighborhood image property value and a difference between light intensity values for the first position and for the second position; and determines a first filtered image property value for the first position as a combination of a first image property value at the first position in the image property map and the combined neighbor image property value.