3D Anomaly Matching in Environmental Image Databases
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods struggle to efficiently compare and monitor changes in an environment over time, particularly in detecting anomalies such as corrosion under insulation, due to the challenges of accurately identifying and tracking the same anomalous parts across different image datasets.
Innovation Solution
A method involving 3D location, classification, and embedding-based clustering of anomalous image portions from multiple image databases, combined with machine learning models and natural language processing, to accurately identify and track anomalies like corrosion under insulation, using point clouds and surface meshes for precise representation and prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional image processing methods are used to compare images from different times, then the process is simple, but the accuracy of identifying and tracking anomalous parts is poor
Solution Approach 1:
The patent segments images into anomalous image portions using machine vision, separating them from normal portions. This allows focused analysis on anomaly regions while maintaining the ability to track them across time through 3D location clustering, resolving the contradiction between simple processing and accurate anomaly identification.
Solution Approach 2:
The patent transitions from 2D image comparison to 3D spatial location comparison by projecting anomalous portions onto 3D point clouds. This dimensional transformation enables accurate tracking of the same anomaly across different time points even when image perspectives change, significantly improving measurement precision without requiring complex multi-view analysis.
2Measurement precision
If comprehensive analysis of environmental changes is performed, then monitoring accuracy is improved, but processing time increases
Solution Approach 1:
The patent extracts only the anomalous portions from complete images using machine vision, then processes only these extracted regions through 3D location clustering and embedding comparison. This extraction approach maintains monitoring accuracy by focusing on critical anomaly regions while dramatically reducing processing time compared to analyzing entire images.
Solution Approach 2:
The patent transforms image data into multiple parameters including 3D location, classification labels, and embeddings. By comparing changes in these parameters across time rather than processing raw images, the system achieves comprehensive monitoring accuracy with reduced computational complexity and processing time.
3Adaptability or versatility
If images from multiple time points are compared to detect changes, then environmental monitoring capability is improved, but difficulty of matching corresponding parts increases
Solution Approach 1:
The patent introduces 3D point clouds as an intermediary representation that bridges images from different time points. By projecting anomalous portions onto a common 3D spatial framework, the system easily matches corresponding parts across time even when camera perspectives or environmental conditions change, significantly reducing matching difficulty while maintaining comprehensive monitoring capability.
Solution Approach 2:
The patent creates a universal 3D location and embedding representation that works across all images from different time points. This universal representation system enables the same processing pipeline to handle diverse imaging conditions and accurately identify corresponding anomalous parts, improving adaptability without increasing matching difficulty.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for processing images of an environment in a first image database from a first time and a second image database from a second time to identify images in each database depicting the same anomalous part of an environment. Image databases and point clouds depicting the environment are accessed along with camera pose information. A first plurality of anomalous image portions are obtained from the first image database and a second plurality of anomalous image portions from the second image database using machine vision. For each anomalous image portion, a 3-dimensional location in the environment, classification and image embedding are obtained. Image portions in the first plurality of image portions and second plurality of image portions are clustered into a first plurality of clusters and second plurality of clusters respectively based on 3D location, classification and embedding. Each cluster corresponds to an anomalous part of the environment. It is determined whether a first cluster from the first plurality of clusters corresponds to the same part of the environment as a second cluster from the second plurality of clusters based on 3D location, classification and embedding of the image portions in the first and second clusters.