Sample Image Processing with Synthetic Negatives for Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of training AI models for anomaly detection in scenarios with scarce samples is exacerbated by the need for large quantities of labeled data, which is costly and labor-intensive to obtain, leading to overfitting and reduced detection accuracy.
Innovation Solution
A method to generate forged negative sample images based on positive sample images and anomaly labeling mask maps, enabling the creation of sufficient datasets for training AI models, thereby improving detection accuracy and resolving overfitting issues.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling is used to obtain labeled data, then detection accuracy can be improved, but labor costs and time consumption increase significantly
Solution Approach 1:
The patent uses a pre-trained anomaly detection model to automatically generate labels for training data by processing positive sample images. This copying approach replicates the labeling function of manual inspection through automated model inference, eliminating the need for time-consuming manual labeling while maintaining label quality consistent with model capabilities
Solution Approach 2:
The system performs self-labeling by using its own anomaly detection model to generate labels for training data. The model processes positive sample images and automatically produces anomaly masks and labels, enabling the system to service its own training data needs without external manual intervention, thus resolving the contradiction between accuracy and time consumption
2Adaptability or versatility
If more labeled data is collected for training, then model generalization capability is improved, but the cost and complexity of data preparation increase
Solution Approach 1:
The patent generates synthetic training data by copying and transforming existing positive sample images through the anomaly detection model. Instead of collecting diverse real-world samples manually, the system creates multiple training variants by processing existing samples through the model, reducing data preparation complexity while maintaining generalization capability
Solution Approach 2:
The system performs preliminary data preparation by using the anomaly detection model to pre-process positive sample images and generate corresponding labels before actual training begins. This preliminary action creates a ready-to-use training dataset with consistent labeling, eliminating the need for complex ongoing data collection and preparation processes
3Reliability
If AI models are trained with sufficient labeled data, then overfitting is reduced, but the resource requirements for data collection and processing increase
Solution Approach 1:
The patent creates synthetic training data by copying existing positive samples through model processing rather than collecting additional real-world data. This approach generates sufficient training variants to prevent overfitting while consuming computational resources only for the synthetic data generation process, avoiding the higher resource costs of manual data collection, storage, and management
Solution Approach 2:
The anomaly detection model serves as an intermediary that transforms positive sample images into labeled training data. This intermediary process efficiently generates training data with consistent quality and labeling, reducing the overall computational resources needed compared to manual data collection and processing workflows
Data Source
AI summary
A sample processing method includes: obtaining a positive sample image; obtaining, based on the positive sample image, an anomaly labeling mask map corresponding to the positive sample image; and generating a forged negative sample image based on the positive sample image and the anomaly labeling mask map, where an anomaly region in the forged negative sample image corresponds to an anomaly labeling region in the anomaly labeling mask map. In this way, a large quantity of forged negative sample images can be automatically generated, to provide sufficient sample datasets for task detection in a scenario with scarce samples.