Stroke-Aware Max-Min Pooling for LSTM OCR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Recurrent neural networks (RNNs), specifically long short-term memory (LSTM) networks, are computationally complex and difficult to deploy in resource-constrained environments for real-time OCR processing, and existing methods to reduce complexity often compromise recognition accuracy, especially when dealing with varied fonts.
Innovation Solution
A text image processing method that applies stroke-aware max-min pooling to pre-process text images before inputting them into an LSTM network, reducing computational complexity and improving recognition accuracy by adapting pooling operations based on stroke thickness, using either max or minimum pooling depending on the thickness relative to an average stroke thickness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If LSTM networks are used for OCR processing, then recognition accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent segments the text image processing into distinct stages: stroke detection, stroke thickness calculation, and conditional pooling operations. By dividing the complex LSTM processing into manageable segments with specialized operations for different stroke types, the computational complexity is reduced while maintaining recognition accuracy across varied fonts.
Solution Approach 2:
The patent applies different pooling operations (max pooling vs. minimum pooling) based on local characteristics of text regions - specifically whether stroke thickness exceeds the average. This local adaptation allows the system to optimize processing for each region's specific features, improving accuracy without uniformly increasing complexity across the entire image.
2Productivity
If parameter reduction techniques are applied to RNN, then computational cost is reduced, but recognition accuracy degrades
Solution Approach 1:
The patent changes the parameter being processed by applying pooling operations that modify pixel values based on stroke thickness characteristics. By transforming the input data representation through conditional pooling before LSTM processing, the system maintains information quality while reducing the dimensional parameters that LSTM must process, thereby improving efficiency without sacrificing accuracy.
3Adaptability or versatility
If a generalized model is created to handle multiple fonts, then adaptability is improved, but model complexity increases
Solution Approach 1:
The patent introduces dynamic adaptability by calculating stroke thickness for each text line and conditionally applying different pooling operations based on whether the stroke thickness exceeds the average. This dynamic adjustment allows the same model structure to adapt to different font characteristics without requiring separate models for each font type, managing complexity through runtime parameter adjustment rather than structural complexity.
Data Source
AI summary
In an optical character recognition (OCR) method for digitizing printed text images using a long-short term memory (LSTM) network, text images are pre-processed using a stroke-aware max-min pooling method before being fed into the network, for both network training and OCR prediction. During training, an average stroke thickness is computed from the training dataset. Stroke-aware max-min pooling is applied to each text line image, where minimum pooling is applied if the stroke thickness of the line is greater than the average stroke thickness, while max pooling is applied if the stroke thickness is less than or equal to the average stroke thickness. The pooled images are used for network training. During prediction, stroke-aware max-min pooling is applied to each input text line image, and the pooled image is fed to the trained LSTM network to perform character recognition.