Learning Device Data Segmentation for Position-Aware Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection devices face challenges in accurately determining anomalies in large datasets due to the lack of management of data piece positions, leading to erroneous determinations and prolonged learning times.
Innovation Solution
A learning device that divides learning target data into pieces, adds identification information for each piece, and generates a learned model using correspondence information to accurately determine anomalies, enabling efficient and accurate machine learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If learning data is divided into pieces to shorten learning time, then learning speed is improved, but position management is lost leading to erroneous anomaly detection
Solution Approach 1:
The patent divides large learning datasets into smaller manageable pieces (patches) to accelerate processing speed. Each patch is independently processed by the neural network, enabling parallel computation and reducing overall learning time while maintaining anomaly detection capability through systematic reconstruction of full-size images from patch-level predictions.
Solution Approach 2:
The patent introduces position information as an intermediary element that bridges the gap between divided patches and the original full-size image structure. This position information is embedded alongside image data and fed to the neural network, enabling the model to understand spatial relationships and correctly reconstruct anomalies in their original positions, thus preventing erroneous detection.
2Measurement precision
If full-size learning data is used to maintain detection accuracy, then anomaly detection precision is improved, but learning time increases significantly
Solution Approach 1:
The patent segments full-size learning images into smaller patches that can be processed more quickly by the neural network. This segmentation reduces the computational burden per training iteration while preserving the essential visual features needed for anomaly detection, thereby shortening learning time without sacrificing detection accuracy.
Solution Approach 2:
The patent adds a new dimension to the input data by incorporating position information alongside image pixel data. This dimensional enhancement allows the neural network to process smaller patches efficiently while still understanding their spatial context within the full-size image, achieving a balance between processing speed and detection accuracy.
3Productivity
If divisional learning data pieces are processed independently, then processing efficiency is improved, but spatial relationship information is lost
Solution Approach 1:
The patent uses position information as an intermediary that carries spatial relationship data from the original full-size image to the divided patches. This position information is concatenated with image data and processed together by the neural network, ensuring that spatial relationships are preserved throughout the independent patch processing workflow.
Solution Approach 2:
The patent merges image data and position information into a unified input tensor that is processed by the neural network. This combination ensures that spatial relationship information is integrated with visual features during processing, allowing the model to maintain understanding of the original image structure even when working with divided patches.
Data Source
AI summary
A learning device includes: a data acquisition unit that acquires learning target data that is full-size data of a learning target; a data generation unit that divides the learning target data to generate multiple pieces of first divided data that is divided data of the learning target data, and adds, to each piece of the first divided data, first identification information for identifying a region of the first divided data in the learning target data; and a model generation unit that generates a learned model for determining an anomaly in the first divided data using first correspondence information that is a set of the first divided data and the first identification information corresponding to the first divided data.