Batch Construction for Text Recognition via Width Thresholding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing OCR technologies face accuracy issues due to the need for padding short text lines to match the length of the longest line in a batch, which can negatively impact the performance of recurrent neural networks like CRNN.
Innovation Solution
The proposed solution involves sorting bounding boxes by width and determining whether to include them in a batch based on the width difference and a threshold value, thereby minimizing the need for padding and improving text recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If text lines are grouped into batches for parallel processing, then processing speed is improved, but short lines must be padded to match the longest line length, which degrades recognition accuracy
Solution Approach 1:
The patent changes the parameter of batch composition by sorting bounding boxes based on width similarity and using a threshold to determine batch membership. This dynamic parameter adjustment allows the system to create batches with minimal width variation, thereby reducing padding requirements while maintaining efficient batch processing for improved productivity without sacrificing recognition accuracy.
2Ease of operation
If short text lines are padded to match the longest line length in a batch, then batch processing can proceed uniformly, but the padding introduces incorrect information that hurts text recognition accuracy
Solution Approach 1:
The system dynamically adjusts batch composition parameters by sorting bounding boxes according to width similarity and applying a threshold criterion. This ensures that batches contain lines of comparable lengths, minimizing the padding required to achieve uniformity while preserving the ease of batch processing operation.
Solution Approach 2:
The patent segments the text lines into multiple batches based on width similarity rather than processing all lines together. By dividing the input into homogeneous subgroups, each batch requires minimal padding, thus maintaining processing uniformity while reducing the harmful effects of padding on recognition accuracy.
3Manufacturing precision
If all text lines are processed individually instead of in batches, then no padding is needed and accuracy is maintained, but the overall processing speed decreases
Solution Approach 1:
The patent optimizes the parameter of batch size and composition by sorting bounding boxes and using width-based thresholding to create optimally sized batches. This allows the system to process multiple lines in parallel (improving productivity) while keeping padding minimal (preserving accuracy), thus resolving the contradiction between individual and batch processing.
Data Source
AI summary
Techniques for constructing padding-sensitive batches for text recognition are provided. In one technique, a first bounding box from a list of bounding boxes is added or included into a batch of zero or more bounding boxes. Each bounding box in the list of bounding boxes surrounds different detected text in a digital image. A second bounding box is identified from the list. The second bounding box is wider than the first bounding box. A difference between (1) a width of the second bounding box and (2) a particular width is determined. The particular width is based on a width of a bounding box in the batch. Based on the difference and a threshold value, it is determined whether to include the second bounding box in the batch. The batch is then input into a test recognition model.