Research on a Traffic Sign Recognition Method under Small Sample Conditions.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Traffic sign recognition systems require large datasets for high accuracy.
• Manual data acquisition and labeling for traffic signs are resource-intensive.
• Few-shot object learning (FSOL) offers a solution to data scarcity challenges.

Purpose of the Study:

• To develop an efficient traffic sign recognition method using FSOL.
• To improve detection accuracy while minimizing the need for extensive training samples.
• To address the limitations of current few-shot object detection algorithms.

Main Methods:

• Modified backbone network with dropout to enhance detection and reduce overfitting.
• Improved region proposal network (RPN) with an attention mechanism for accurate candidate box generation.
• Feature pyramid network (FPN) for multi-scale feature extraction and fusion.

Main Results:

• Achieved a 4.27% improvement on 5-way 3-shot tasks and 1.64% on 5-way 5-shot tasks compared to the baseline.
• Demonstrated superior performance over current few-shot object detection algorithms on the PASCAL VOC dataset.
• The proposed method effectively handles the challenge of limited training data in traffic sign recognition.

Conclusions:

• The proposed FSOL-based method significantly enhances traffic sign recognition accuracy.
• The integration of attention mechanisms and FPN improves feature representation and detection capabilities.
• This approach provides a viable solution for developing robust traffic sign recognition systems with limited data.