3D Partial-Data Learning for Rare-Anomaly Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for anomaly detection in objects using machine learning struggle with insufficient teaching data when anomalies are rare, leading to reduced model accuracy.
Innovation Solution
An object evaluation apparatus and method that utilizes an evaluation model generated from partial three-dimensional data, where shapes of at least two pieces of data are identical within a predetermined error, allowing for high-accuracy anomaly detection without pre-prepared teaching data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If symmetry of an object is used for increasing the number of pieces of teaching data, then the number of teaching data is increased, but the number of pieces of teaching data with anomaly remains insufficient
Solution Approach 1:
The patent divides three-dimensional data of an object into multiple pieces of partial data. By segmenting the data in this way, the system can create multiple learning datasets from a single object, effectively increasing the quantity of teaching data available for training the anomaly detection model, while ensuring that each segment contains sufficient information for reliable anomaly detection
Solution Approach 2:
The patent transitions from using two-dimensional images to utilizing three-dimensional data for object analysis. This dimensional change enables more effective data segmentation and allows the system to generate multiple partial data pieces from a single three-dimensional object, thereby increasing teaching data quantity while maintaining anomaly detection accuracy
Data Source
AI summary
A model acquisition unit acquires an evaluation model which is generated by setting, as learning data, some of a plurality of pieces of partial data generated by dividing three-dimensional data indicating a shape of an object into a plurality of pieces. An evaluation data generation unit generates, by using at least some of rest of the plurality of pieces of partial data as input data to evaluation model, evaluation data for evaluating whether the object has an anomaly. Three-dimensional shapes respectively indicated by at least two pieces of the partial data are identical within a range including a predetermined error. At least one of pieces of partial data of which three-dimensional shapes are identical with each other is included in first partial data for generating the learning data, and at least one other of pieces of partial data is included in second partial data to be the input data.