Vertical Text OCR Using Rotation and Ensemble Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OCR systems struggle to accurately recognize vertically oriented alphanumeric text due to the scarcity of such text instances in real-world scenarios, leading to inadequate performance when confronted with vertical text layouts.
Innovation Solution
A system and method that employs a trained text detector to identify regions-of-interest in images, rotates the detected text by 90 degrees, and uses a trained ensemble of two different text recognition models to enhance accuracy and reliability in recognizing vertically oriented alphanumeric text.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing OCR systems are used for vertical text recognition, then horizontal text recognition performance is maintained, but vertical text recognition accuracy deteriorates
Solution Approach 1:
The system segments the text recognition task by detecting text orientation first, then applying different recognition models for horizontal versus vertical text. This segmentation allows each model to specialize in its optimal orientation, resolving the contradiction between maintaining horizontal text performance and improving vertical text accuracy.
Solution Approach 2:
Instead of attempting to make a single OCR model handle both orientations equally, the system inverts the approach by creating separate specialized models for horizontal and vertical text. This inversion resolves the contradiction by accepting that different models are needed for different orientations rather than forcing one model to be universally competent.
2Measurement precision
If OCR models are trained on standard datasets like IC13 and IC15, then general text recognition capability is improved, but vertical text recognition capability deteriorates
Solution Approach 1:
The system applies local quality by training different recognition models with different characteristics suited to specific text orientations. The vertical text recognition model is specifically trained on datasets containing vertical text samples, giving it local expertise in vertical text patterns rather than relying on general horizontal text training data.
Solution Approach 2:
The system changes the training parameters and data composition for the vertical text model specifically. By incorporating vertical text samples from sources like license plates and signage into the training dataset, the model's parameters adapt to vertical text patterns, resolving the contradiction between general text recognition capability and vertical text specialization.
3Measurement precision
If a single text recognition model is used, then system complexity is reduced, but recognition accuracy for vertical text deteriorates
Solution Approach 1:
The system implements dynamics by making the model selection adaptive rather than static. The orientation detection component dynamically determines which recognition model to apply based on the detected text orientation in each image or region, allowing the system to use multiple specialized models only when needed while maintaining simplicity for standard cases.
Data Source
AI summary
A system for recognizing vertically oriented alphanumeric text in images, the system including a processor configured to receive one or more images comprising vertically oriented alphanumeric text and detect one or more regions-of-interest in each image via a trained text detector. The processor is configured to execute a cropping of the detected one or more regions-of-interest encompassing vertically oriented alphanumeric text from each image to obtain one or more text crop portions and rotate the one or more text crop portions to obtain one or more orthogonally rotated text crop portions. The processor is configured to execute a trained ensemble of two different text recognition models on each of the obtained one or more text crop portions and the one or more orthogonally rotated text crop portions and generate a set of candidate recognized text strings based on the executed trained ensemble and determine a final recognized text string.