Lightweight traffic sign detection algorithm with noise suppression and semantic enhancement.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Edge computing deployment of traffic sign detection algorithms faces challenges with high computational overhead.
• Existing methods struggle with detecting small or deformed traffic signs effectively.

Purpose of the Study:

• To develop a lightweight and high-precision traffic sign detection algorithm for edge devices.
• To improve the detection capabilities for small and deformed traffic signs.

Main Methods:

• Proposed LNSE-YOLO algorithm based on YOLOv11, incorporating Noise Suppression and Semantic Enhancement (NSSE) module.
• Integrated Edge-Driven Feature Enhancement Network (ED-FEN) and Local Deformable Attention (LDA) module.
• Employed channel pruning strategy for model lightweighting.

Main Results:

• The high-precision NSE-YOLO model improved mAP@50 by 5.5 and 2.5 percentage points on TT100K and CCTSDB datasets, respectively.
• The lightweight LNSE-YOLO model maintained accuracy while significantly reducing parameters.
• Achieved comparable computational cost to the baseline YOLOv11n.

Conclusions:

• The proposed LNSE-YOLO algorithm effectively addresses limitations of existing traffic sign detection on edge devices.
• The methodology demonstrates practical utility and significant improvements in accuracy and efficiency.
• LNSE-YOLO offers a viable solution for real-world intelligent transportation systems.