Edge OCR Classification Head Updates for Real-Time Online Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional OCR systems struggle with real-time learning and parameter updates in complex scenarios due to their inability to dynamically adapt to new input data, especially in edge devices, leading to reduced accuracy and efficiency.
Innovation Solution
A method and apparatus for training a character recognition model on edge devices using a state space model to update the classification head parameters in real-time based on inconsistencies between predicted and labeled strings, reducing data processing and simplifying the training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If traditional OCR systems use predefined template matching or statistical learning methods, then real-time processing requirements can be met, but accuracy and efficiency are reduced in complex scenarios
Solution Approach 1:
The patent implements dynamic adaptability by enabling the OCR model to continuously learn and update its parameters in real-time on edge devices. The system transitions from static predefined templates to dynamic neural network models that can adapt to new characters, fonts, and environmental conditions while maintaining real-time processing capabilities through optimized inference engines.
Solution Approach 2:
The patent changes the fundamental parameters of the OCR system by replacing traditional template matching algorithms with deep learning models. This parameter change enables the system to handle complex scenarios with varying illumination, angles, and distortions while maintaining real-time performance through hardware-accelerated inference.
2Measurement precision
If deep learning models are deployed on edge devices for high-precision recognition, then accuracy improves in complex scenarios, but the models are unable to achieve real-time online learning
Solution Approach 1:
The patent makes the edge device OCR system dynamic by implementing continuous online learning capabilities. The model can update its weights and biases in real-time using new input data, allowing it to adapt to changing environments and new character types while deployed on resource-constrained edge devices.
Solution Approach 2:
The patent enables the OCR model to serve itself by implementing self-updating mechanisms on edge devices. The system can automatically learn from new data, update its parameters, and improve its performance without requiring external retraining or manual intervention, achieving autonomous adaptation in real-time.
3Quantity of substance
If lightweight deep OCR models are designed for edge devices, then device resource requirements are reduced, but real-time online learning and dynamic parameter updates are unable to be achieved
Solution Approach 1:
The patent segments the model updating process into efficient, incremental updates that can be performed on edge devices with limited resources. By dividing the learning process into small batches and using optimized update algorithms, the system achieves dynamic parameter updates without requiring large memory resources or extensive computational power.
Solution Approach 2:
The patent implements partial updates by selectively updating only the necessary model parameters based on new data, rather than retraining the entire model. This partial action approach reduces computational overhead and memory requirements while still achieving effective adaptation to new scenarios on resource-constrained edge devices.
Data Source
AI summary
Described is a method and apparatus for training a character recognition model, a computer device, and a storage medium. The method includes: acquiring an input image and a labeled string of the input image; performing character recognition on the input image via the character recognition model pre-deployed on an edge device to obtain a predicted string of the input image; and performing a parameter update on a classification head in the character recognition model via a state space model in a case where the predicted string is inconsistent with the labeled string; wherein the state space model contains a state equation and an observation equation, the state equation is used to indicate an evolutionary relationship of a classification head parameter between different time steps, and the observation equation is used to generate an observable observation character based on the classification head parameter.