Document Image Anomaly Detection Using ROI-Specific Transforms
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital identity verification systems face challenges in detecting document anomalies due to high variability across document types and countries, noise levels, lack of genuine reference documents, limited fraudulent data, and performance degradation with new fraud types, requiring a modular, adaptable, scalable, interpretable, and customized solution.
Innovation Solution
A method involving segmentation of document images into regions of interest (ROIs), applying transformations to generate transform-specific and region-specific features, and computing anomaly scores with modular thresholds to detect anomalies, allowing for local adaptability and interpretability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a comprehensive fraud detection model is trained across thousands of different document classes, then the model can detect various fraud types, but the model performance deteriorates when new fraud types are introduced and global model updates are required
Solution Approach 1:
The patent segments the document detection task into separate document class models, each specialized for a specific document type (e.g., passports, driver's licenses). This allows each model to be optimized for its specific document class without being diluted by the diversity of other classes, thereby maintaining high reliability when detecting new fraud types in a given document class without requiring global model updates.
2Measurement precision
If fraud-specific anomaly detection models are trained for specific fraud types, then detection performance improves for known frauds, but the models fail to detect unknown fraud types
Solution Approach 1:
Instead of training models to recognize specific fraud types (supervised approach), the patent inverts the approach by training models to recognize only legitimate documents (one-class classification). The anomaly detection is performed by identifying deviations from the learned legitimate pattern, which allows the system to detect any unknown fraud type without requiring specific training data for each fraud type.
3Adaptability or versatility
If general fraud detection models are used, then coverage across document classes is achieved, but explanations for decisions are not provided
Solution Approach 1:
The patent applies local quality by providing detailed explanations at the specific region level where anomalies were detected. Instead of providing only a general decision, the system identifies and explains the specific ROI (region of interest) and the specific transformation features that indicated the anomaly, giving both location and reason for the detection decision.
4Measurement precision
If multiple transformations are applied to document images, then feature extraction is improved, but processing time increases
Solution Approach 1:
The patent applies dynamics by making the number and type of transformations applied adaptive rather than fixed. The system dynamically selects which transformations to apply based on the specific document class, the detected anomaly type, and the ROI being analyzed. This allows high feature extraction quality when needed while reducing processing time for simpler cases or when fewer transformations are required.
Data Source
AI summary
Described are methods and systems for training a system for detecting anomalies in images of documents in a class of documents. A plurality of training document images of training documents in a class of documents are obtained. For each training document image, the training document image is segmented into a plurality of region of interest (ROI) images, each ROI image corresponding to a respective ROI of the training document. For each ROI image, a plurality of transformations are applied to the ROI image to generate respective transform-specific features for the ROI image and respective transform-specific anomaly scores from the transform-specific features. Based on the respective anomaly scores of the plurality of training document images, a transform-specific threshold is computed for each transformation to separate document images containing an anomaly from document images not containing an anomaly.