Character-Wise OCR Training With Contrastive Learning and Location Tokens
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OCR systems face challenges with data augmentation techniques that can lead to character loss or alteration, and loss functions are not suitable for improving character recognition accuracy.
Innovation Solution
A character-wise supervised contrastive learning method using a deep learning model that generates synthetic images with varying character appearances and incorporates location information for training, enabling effective feature extraction and recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation techniques are applied to increase training data diversity, then the amount of training data increases, but character loss or alteration occurs
Solution Approach 1:
The patent segments the training data into character-level units with explicit location information (bounding boxes). By processing images at the character level rather than applying global augmentation transformations, the system can generate diverse training samples without losing or altering character integrity. Each character is independently identified and labeled with its position, allowing precise control over data augmentation while maintaining character reliability.
Solution Approach 2:
The patent introduces location information (bounding boxes) as an intermediary element that mediates between the original image and the augmented training data. This intermediary provides precise spatial information about each character, enabling the system to apply data augmentation techniques while tracking and preserving character positions and integrity throughout the transformation process.
2Productivity
If conventional loss functions are used for training, then the training process is simple, but character recognition accuracy does not improve
Solution Approach 1:
The patent changes the loss function parameters from conventional approaches to a character-wise supervised contrastive loss function. This new loss function incorporates location information and character-level supervision signals, fundamentally altering the training objective to simultaneously optimize for character recognition accuracy and spatial positioning. The modified loss function maintains training efficiency while significantly improving measurement precision in character recognition.
3Ease of manufacture
If real images without location information are used for training, then data collection is easier, but the model cannot learn spatial positioning of characters
Solution Approach 1:
The patent applies preliminary action by generating synthetic training images with embedded location information before actual training begins. Using a text rendering engine, the system pre-generates images with known character positions and bounding boxes, creating a foundation of labeled training data that enables the model to learn spatial positioning from the outset, even when real images lack such information.
Solution Approach 2:
The patent creates synthetic copies of text images with known location information using a text rendering engine. These synthetic copies serve as proxy training data that replicates the characteristics of real images while providing the necessary location annotations. The model learns from these copied synthetic images and generalizes to real images, effectively transferring the location information learning capability without requiring manually annotated real image data.
Data Source
AI summary
An OCR method using a character-wise supervised contrastive learning model includes receiving an input image; extracting, from the input image, a token sequence representing character information and location information of the input image by means of a character-wise supervised contrastive learning model; and converting the token sequence into visualized information.