Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Lightweight Multi-Module Fusion Approach for Korean Character Recognition

Inho Jake Park, Jaehoon Jay Jeong, Ho-Sang Jo

TL;DR

SDA-Net targets robust single-character Korean OCR by combining Stroke-Sensitive Attention with edge-aware spatial cues, a Dynamic Context Encoding module, and a U‑Net–inspired skip fusion in a lightweight ResNet-based backbone. The model achieves high accuracy in challenging conditions while maintaining low memory and compute demands (5.6M parameters, 3.4 GFLOPs), enabling real-time edge deployment. A novel ACPSLD dataset for license-plate–level single-character recognition supports real-world evaluation, including strict KNPA-aligned correctness metrics and on-site testing with notable performance. Ablation studies demonstrate the incremental gains from SSA, edge attention, and DCE, underscoring the value of integrated attention, context refinement, and multi-scale fusion for robust OCR. The work advances practical OCR for traffic surveillance by delivering accurate, efficient recognition under diverse, real-world conditions.

Abstract

Optical Character Recognition (OCR) is essential in applications such as document processing, license plate recognition, and intelligent surveillance. However, existing OCR models often underperform in real-world scenarios due to irregular text layouts, poor image quality, character variability, and high computational costs. This paper introduces SDA-Net (Stroke-Sensitive Attention and Dynamic Context Encoding Network), a lightweight and efficient architecture designed for robust single-character recognition. SDA-Net incorporates: (1) a Dual Attention Mechanism to enhance stroke-level and spatial feature extraction; (2) a Dynamic Context Encoding module that adaptively refines semantic information using a learnable gating mechanism; (3) a U-Net-inspired Feature Fusion Strategy for combining low-level and high-level features; and (4) a highly optimized lightweight backbone that reduces memory and computational demands. Experimental results show that SDA-Net achieves state-of-the-art accuracy on challenging OCR benchmarks, with significantly faster inference, making it well-suited for deployment in real-time and edge-based OCR systems.

A Lightweight Multi-Module Fusion Approach for Korean Character Recognition

TL;DR

SDA-Net targets robust single-character Korean OCR by combining Stroke-Sensitive Attention with edge-aware spatial cues, a Dynamic Context Encoding module, and a U‑Net–inspired skip fusion in a lightweight ResNet-based backbone. The model achieves high accuracy in challenging conditions while maintaining low memory and compute demands (5.6M parameters, 3.4 GFLOPs), enabling real-time edge deployment. A novel ACPSLD dataset for license-plate–level single-character recognition supports real-world evaluation, including strict KNPA-aligned correctness metrics and on-site testing with notable performance. Ablation studies demonstrate the incremental gains from SSA, edge attention, and DCE, underscoring the value of integrated attention, context refinement, and multi-scale fusion for robust OCR. The work advances practical OCR for traffic surveillance by delivering accurate, efficient recognition under diverse, real-world conditions.

Abstract

Optical Character Recognition (OCR) is essential in applications such as document processing, license plate recognition, and intelligent surveillance. However, existing OCR models often underperform in real-world scenarios due to irregular text layouts, poor image quality, character variability, and high computational costs. This paper introduces SDA-Net (Stroke-Sensitive Attention and Dynamic Context Encoding Network), a lightweight and efficient architecture designed for robust single-character recognition. SDA-Net incorporates: (1) a Dual Attention Mechanism to enhance stroke-level and spatial feature extraction; (2) a Dynamic Context Encoding module that adaptively refines semantic information using a learnable gating mechanism; (3) a U-Net-inspired Feature Fusion Strategy for combining low-level and high-level features; and (4) a highly optimized lightweight backbone that reduces memory and computational demands. Experimental results show that SDA-Net achieves state-of-the-art accuracy on challenging OCR benchmarks, with significantly faster inference, making it well-suited for deployment in real-time and edge-based OCR systems.

Paper Structure

This paper contains 39 sections, 18 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Key Contributions
  3. Related Works
  4. Attention-Based OCR Models
  5. Lightweight OCR Models
  6. Proposed SDA-Net
  7. Method
  8. Overall Architecture
  9. Feature Extraction
  10. Dual Attention Mechanism
  11. Channel Attention
  12. Spatial Attention (Edge Attention)
  13. Dynamic Context Encoding
  14. Feature Fusion with Skip Connection
  15. Prediction
  16. ...and 24 more sections

Figures (5)

  • Figure 1: Proposed SDA-Net architecture
  • Figure 2: Example license plate types and formats in the ACPSLD dataset.
  • Figure 3: Evaluation process for license plate recognition using real-time CCTV feed.
  • Figure 4: Various license plates from on-site locations
  • Figure 5: Examples of difficult cases in the ACPSLD dataset, including dust-covered plates, occluded characters, and motorcycle license plates.