LTH-Pruned Data Models for Real-Time Document Text Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document processing techniques for image-based text extraction are computationally complex and have large memory footprints, making them unsuitable for real-time applications.
Innovation Solution
A method and system that utilize the Lottery Ticket Hypothesis (LTH) algorithm to prune and trim a pre-trained base model, followed by a Knowledge Distillation algorithm to generate a structured pruned data model for text detection, involving filter sparsity thresholding and zero percentage-based filter discarding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If state of the art techniques are used for document processing and text extraction, then text detection accuracy is maintained, but computational load and memory footprint increase significantly
Solution Approach 1:
The patent segments the neural network model into distinct components (backbone, neck, head) and applies different pruning strategies to each segment. The backbone uses LTH-based unstructured pruning while the neck and head use structured channel pruning, allowing targeted optimization that reduces computational complexity while preserving detection accuracy.
Solution Approach 2:
The patent changes the sparsity parameter from unstructured (random) to structured (channel-based) pruning. By transforming the pruning approach from random weight removal to structured channel removal, the model achieves better hardware compatibility and reduced computational load while maintaining accuracy through knowledge distillation.
2Measurement precision
If state of the art techniques are used for document processing, then text detection capability is achieved, but processing time increases making it unsuitable for real-time applications
Solution Approach 1:
The patent performs preliminary pruning of the neural network model before deployment using the LTH algorithm. By pre-identifying and removing redundant channels and filters in advance, the model achieves faster inference time during real-time processing while maintaining detection capability through the preserved important features.
Solution Approach 2:
The patent uses knowledge distillation to copy knowledge from the large pre-trained model to a smaller pruned model. The teacher model (original large model) transfers its learned patterns and features to the student model (pruned model), enabling the smaller model to achieve real-time processing speeds while maintaining detection accuracy.
Data Source
AI summary
State of the art techniques used for document processing and particularly for handling processing of images for data extraction have the disadvantage that they have large computational load and memory footprint. The disclosure herein generally relates to text processing, and, more particularly, to a method and system for generating a data model for text extraction from documents. The system prunes a pretrained base model using a Lottery Ticket Hypothesis (LTH) algorithm, to generate a LTH pruned data model. The system further trims the LTH pruned data model to obtain a structured pruned data model, which involves discarding filters that have filter sparsity exceeding a threshold of filter sparsity. The structured pruned data model is then trained from a teacher model in a Knowledge Distillation algorithm, wherein a resultant data model obtained after training the structured pruned data model forms the data model for text detection.