Object Detection in Unstructured Documents Using Synthetic Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for object detection in digitized documents, such as signatures, are inefficient due to the lack of customizable training datasets and the inability to detect objects in unstructured documents where positions are undefined, leading to inaccurate and slow detection.
Innovation Solution
A system and method using a neural network-based model trained with a generated training dataset, created by blending pre-defined object images with pruned portions of original documents through wavelet transformation, enabling detection of objects like signatures in unstructured documents.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning neural networks are used for object detection, then detection accuracy may be improved, but significant training data is required which is challenging to obtain
Solution Approach 1:
The patent creates synthetic training data by copying and blending signature images with document backgrounds. Instead of requiring extensive real-world annotated data, the system generates artificial training examples by superimposing signature images onto document templates, thereby reducing the need for large quantities of actual training data while maintaining detection accuracy
Solution Approach 2:
The system uses available document templates and signature samples to automatically generate its own training dataset. Rather than relying on external data sources or manual annotation efforts, the system serves itself by synthesizing training data from its own operational resources, eliminating the bottleneck of data collection and annotation
2Ease of manufacture
If conventional computer vision methods are used for signature extraction, then the process is simple, but the method fails when other large contours are present on the image
Solution Approach 1:
The patent transforms the detection approach by changing from contour size-based filtering to neural network-based pattern recognition. Instead of relying on the parameter of contour area to identify signatures, the system uses trained neural networks that can distinguish signatures from other large contours (like logos or text) based on learned visual patterns, thereby maintaining simplicity while improving reliability
Solution Approach 2:
The patent replaces the mechanical contour detection and filtering system with an intelligent neural network-based recognition system. The traditional mechanical approach of removing small contours assumes signatures are always large, but the neural network substitution allows the system to reliably identify signatures regardless of their size or the presence of other large elements, maintaining ease of use while dramatically improving reliability
3Adaptability or versatility
If training datasets are customized for diverse object locations, then detection coverage improves, but dataset creation becomes complex and time-consuming
Solution Approach 1:
The patent implements dynamic dataset generation where training examples are created by randomly placing signatures at various positions on document templates during the training process. This dynamic approach allows the model to learn to detect signatures anywhere on the document without requiring separate static datasets for each position, thereby improving detection coverage while keeping the dataset creation process simple and automated
Solution Approach 2:
The system performs preliminary synthesis of diverse training scenarios by pre-generating images with signatures at various locations, orientations, and scales before training begins. This preliminary action creates a comprehensive training corpus that covers all possible detection scenarios, allowing the model to be versatile without requiring complex manual dataset curation during deployment
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The system allows for accurate and real-time detection of objects in unstructured documents, overcoming the limitations of conventional techniques by using diverse and customizable training datasets.
Implementation Method 1
The dataset generator blends the first input image with the pruned image using a blending technique which comprises a wavelet transformation technique
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
Systems and methods for facilitating an automated detection of an object in a test document are disclosed. A system may include a processor including a dataset generator. The dataset generator may obtain a first input image and a first original document from a data lake. The dataset generator may prune a portion of the first original document to obtain a pruned image. The dataset generator may blend the first input image with the pruned image to generate a modified image. The modified image may include the pruned image bearing the first pre-defined representation. The modified image may be combined with the first original document to generate a training dataset. The training dataset may be utilized to train a neural network based model to obtain a trained model for the automated detection of the object in the test document.