Robust Semi-Supervised Traffic Sign Recognition via Self-Training and Weakly-Supervised Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Traffic sign recognition faces challenges due to the increasing volume of unlabeled data and class imbalance.
• Traditional methods require extensive labeled data, which is costly and time-consuming to acquire.
• Deep neural networks demand large datasets, exacerbating the need for efficient data utilization.

Purpose of the Study:

• To develop a semi-supervised classification technique robust to small and imbalanced datasets for traffic sign recognition.
• To effectively leverage large amounts of unlabeled data alongside limited labeled samples.
• To enhance the performance and efficiency of traffic sign classifiers.

Main Methods:

• Proposed a novel semi-supervised classification framework integrating weakly-supervised learning, self-training, and self-paced learning.
• Developed an attention map generation method to augment the training set.
• Implemented a pseudo-label generation and selection algorithm for unlabeled data utilization.

Main Results:

• The method normalizes class-wise confidence levels to address data imbalance and hard-to-learn samples.
• Jointly learned a model and optimized pseudo-labels from unlabeled data.
• Successfully enlarged the training set, improving deep learning model performance.

Conclusions:

• The proposed technique demonstrates effectiveness on public traffic sign recognition datasets.
• Offers a potential solution for building efficient and high-quality traffic sign classifiers in practical applications.
• Addresses key challenges in traffic sign recognition, including data scarcity and imbalance.