Document Anomaly Detection via ROI Segmentation and Transform-Specific Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Digital identity verification systems face challenges in detecting anomalies in documents due to high variability across document types, intrinsic and extrinsic noise, lack of genuine reference documents, limited fraudulent data, and the need for modular and adaptable solutions that can quickly respond to new fraud types without sacrificing performance or causing delays.
Innovation Solution
A computer-implemented method that segments document images into regions of interest (ROI) and applies various transformations to generate transform-specific and region-specific features, enabling modular and interpretable anomaly detection by computing anomaly scores and thresholds, allowing for continuous updates and efficient adaptation to new document features and fraud types.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a global fraud detection model is trained across thousands of different document classes, then the model can handle diverse document types, but updating the model for new fraud types causes performance degradation and delays
Solution Approach 1:
The patent segments the global fraud detection model into multiple independent document class models. Each model is trained separately on its specific document type, allowing updates to be made locally without affecting other document classes. This segmentation enables the system to adapt to new fraud types in specific document classes while maintaining stable performance in other classes, resolving the contradiction between versatility and reliability.
2Quantity of substance
If comprehensive fraudulent data is collected for training, then the model can detect various fraud types, but fraudulent data is frequently lacking or extremely limited
Solution Approach 1:
The patent uses synthetic data generation to create virtual fraudulent document images that replicate real fraud patterns. These synthetic copies are generated using GANs (Generative Adversarial Networks) and other synthesis techniques, allowing the model to be trained on abundant simulated fraudulent data without needing to collect actual fraudulent documents. This resolves the contradiction by providing unlimited training data while avoiding the difficulties of real data collection.
Solution Approach 2:
The patent employs synthetic data that can be generated on-demand and discarded after use, replacing the need for collecting and storing valuable fraudulent documents. This approach provides unlimited training data at minimal cost, resolving the contradiction between data quantity and collection difficulty.
3Reliability
If general fraud detection models are used, then the system can detect fraud, but the models do not provide explanations about the decision to mark documents as fraudulent
Solution Approach 1:
The patent introduces an explanation module that acts as an intermediary between the fraud detection model and the user. This module analyzes the model's decisions and provides human-readable explanations for why documents were flagged, including which features contributed most to the anomaly score. This intermediary layer preserves both the detection capability and the explanatory information, resolving the contradiction between reliability and information loss.
4Quantity of substance
If the system processes entire document images, then all information is analyzed, but the level of noise in the image reduces detection accuracy
Solution Approach 1:
The patent extracts and processes only the most relevant regions of interest (ROIs) from document images rather than analyzing the entire image. By identifying and focusing on specific areas containing critical information (such as text regions, security features, or biometric data), the system reduces the impact of noise from irrelevant areas while maintaining detection accuracy. This extraction approach resolves the contradiction by processing less data with higher precision.
Data Source
Figure 1
Figure 2
Figure 3
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.