SNoW Classifier Feature Pruning for Low-Memory OCR
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated License Plate Recognition (ALPR) systems face challenges in optical character recognition due to non-uniform backgrounds, noise, and high memory requirements, particularly when dealing with multiple state fonts, which leads to increased RAM utilization and reduced throughput.
Innovation Solution
A method is introduced to optimize SNoW classifiers by identifying and using only the top N Successive Mean Quantization Transform (SMQT) features, reducing the feature set from 512 to 60, and converting weights from doubles to single byte integers, resulting in an 88% reduction in RAM utilization and further memory footprint reduction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a separate OCR engine is trained for each state font to handle individual fonts across different U.S. states, then OCR accuracy is improved, but RAM utilization and classifier size grow rapidly
Solution Approach 1:
The patent implements a universal OCR engine that can handle multiple state fonts simultaneously through a single classifier trained on concatenated feature vectors. Instead of creating separate specialized engines for each font, the system uses a unified model that processes all fonts through shared neural network layers, achieving multi-functionality without proportional increases in memory resources.
Solution Approach 2:
The patent combines multiple font-specific feature vectors into a single concatenated feature vector that serves as input to a unified classifier. By merging the feature representations of different fonts and training a single engine on this combined input, the system reduces the number of separate classifiers needed while maintaining the ability to distinguish between different font types.
2Productivity
If multiple instances of the ALPR engine are launched to achieve high throughput via parallelization, then throughput is improved, but memory management problems are compounded
Solution Approach 1:
The unified classifier serves multiple processing threads and parallel instances simultaneously without requiring separate memory allocations for each instance. The shared model weights and single classifier structure allow multiple engine instances to operate in parallel while sharing the same memory footprint, enabling high throughput through parallelization without compounding memory requirements.
3Measurement precision
If the full set of 512 SMQT features is used in the classifier, then OCR accuracy is maintained, but the memory footprint is large
Solution Approach 1:
The patent extracts and utilizes only the most discriminative features from the full set of 512 SMQT features by concatenating selected feature vectors from different fonts. This selective extraction approach identifies and retains the features that contribute most to accurate classification while discarding redundant features, thereby reducing memory requirements while preserving OCR accuracy.
Data Source
AI summary
Methods and systems for reducing the required footprint of SNoW-based classifiers via optimization of classifier features. A compression technique involves two training cycles. The first cycle proceeds normally and the classifier weights from this cycle are used to rank the Successive Mean Quantization Transform (SMQT) features using several criteria. The top N (out of 512 features) are then chosen and the training cycle is repeated using only the top N features. It has been found that OCR accuracy is maintained using only 60 out of 512 features leading to an 88% reduction in RAM utilization at runtime. This coupled with a packing of the weights from doubles to single byte integers added a further 8× reduction in RAM footprint or a reduction of 68× over the baseline SNoW method.