Signature Pre-processing for Classification Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional signature analysis techniques result in inconsistent feature extraction due to varying signature positions and digitization methods, leading to less accurate signature type classification and incorrect acceptance decisions in form documents.
Innovation Solution
The proposed solution involves pre-processing signature data by aligning and extracting signatures from form documents using anchor points and text extraction libraries, followed by training a signature type classifier using the aligned data to improve classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional signature extraction techniques are applied to historical documents, then signature data can be obtained for training, but the features manifest inconsistently resulting in lower classification accuracy
Solution Approach 1:
The patent applies preliminary action by pre-processing signature data before training the classifier. Specifically, it extracts signatures from historical documents, aligns them to a standard format, and normalizes their features before using them to train the signature type classifier. This preliminary preparation ensures that the training data has consistent features, which directly improves classification accuracy.
Solution Approach 2:
The patent changes parameters of the signature data by transforming extracted signatures into a standardized format with consistent features. It modifies parameters such as signature positioning, scaling, and feature extraction methods to ensure uniformity across all training samples. This parameter standardization resolves the feature inconsistency problem while maintaining the ability to distinguish different signature types.
2Adaptability or versatility
If different signature extraction techniques are applied to historical documents, then various signature formats can be captured, but the features manifest inconsistently leading to incorrect acceptance decisions
Solution Approach 1:
The patent standardizes parameters of extracted signatures by transforming them into a uniform format regardless of their original extraction method. It applies consistent feature extraction, alignment, and normalization processes to all signatures, ensuring that the classifier receives reliable, consistent input data. This parameter standardization maintains adaptability to various signature formats while ensuring reliable acceptance decisions.
3Quantity of substance
If historical documents are included in the corpus in an inconsistent manner, then more training data is available, but the classifier becomes less accurate
Solution Approach 1:
The patent applies preliminary action by systematically pre-processing all training data before classifier training. It extracts signatures from historical documents using consistent methods, aligns them to a standard format, and normalizes their features. This preliminary standardization ensures that the increased training data volume does not degrade accuracy; instead, it improves the classifier's ability to learn from diverse yet consistent examples.
Solution Approach 2:
The patent transforms parameters of training data by converting heterogeneous signature formats into a standardized representation. It applies uniform feature extraction, positioning, and scaling to all training samples, ensuring that the increased data quantity contributes to better generalization rather than introducing noise. This parameter standardization enables the classifier to achieve high accuracy despite the large and diverse training corpus.
Data Source
AI summary
The following generally relates to pre-processing signature data. In some examples, the extracted signature data may be used to train a signature classification model configured to identify a signature type for an input signature. As part of the pre-processing, the techniques disclosed herein relate to extracting the signature data from a corpus of signed documents. For example, techniques disclosed herein may use anchor points to define a boundary box. Additionally, the pre-processing may include aligning the image data of the signatures, for example, by correcting a skew. The extracted and aligned signatures may be used to train the signature classification model.