Scene Text Recognition With Writing Direction and Count Tokens
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional scene character recognition models struggle to accurately recognize both horizontal and vertical writing directions without sufficient teacher data for both writing types, particularly challenging due to the scarcity of vertical writing data in real-world environments.
Innovation Solution
An information processing apparatus that includes a feature extraction unit and a character string estimation unit, which estimates the writing direction and number of characters to enhance character recognition accuracy by sharing model parameters between horizontal and vertical writing, and uses tokens to distinguish writing directions and character counts in an autoregressive decoder.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single model is used to recognize both horizontal and vertical writing, then model complexity is reduced, but recognition accuracy deteriorates due to insufficient writing direction discrimination
Solution Approach 1:
The model is segmented into distinct components: a shared base model for common feature extraction, and separate writing direction recognition models for horizontal and vertical text. This segmentation allows each component to specialize, improving overall accuracy while maintaining reasonable complexity through parameter sharing in the base model.
Solution Approach 2:
A writing direction recognition module acts as an intermediary between the input image and the main recognition model. This intermediary first determines the writing direction and provides this information as guidance to the recognition model, enabling accurate recognition without requiring separate complete models for each direction.
2Measurement precision
If separate models are used for horizontal and vertical writing recognition, then recognition accuracy is improved, but device complexity increases
Solution Approach 1:
The model merges common feature extraction capabilities into a shared base model that serves both horizontal and vertical writing recognition. By combining the shared components with direction-specific modules, the system achieves accurate recognition for both directions while reducing overall complexity through parameter sharing.
3Device complexity
If all model parameters are shared between horizontal and vertical writing, then device complexity is reduced, but recognition accuracy deteriorates due to inability to distinguish writing directions
Solution Approach 1:
The model applies local quality by having different parameter sharing strategies for different parts of the model. The base model uses full parameter sharing for common features, while the writing direction recognition modules use direction-specific parameters to capture directional characteristics. This localized differentiation maintains simplicity where possible while improving accuracy where needed.
4Ease of operation
If writing direction information is not provided to the model, then ease of operation is improved, but recognition accuracy deteriorates due to inability to adapt to different writing directions
Solution Approach 1:
The model implements self-service by automatically performing writing direction recognition and using this information to adapt its processing. The system autonomously determines the writing direction from the input image and configures its recognition parameters accordingly, eliminating the need for manual direction specification while maintaining high accuracy.
Data Source
AI summary
An information processing apparatus includes processing circuitry configured to extract an image feature from a character image, and estimate a character string from a writing direction and the image feature.