學(xué)者網(wǎng)
HM-Net: An improved few-shot learning model for the detection of traffic signs
Highlights
- •
Inspired by hierarchical structures of images, this study designed a Hyperbolic Hierarchical Constraint (HHC) algorithm.
- •
A Hyperbolic Relationship Metric (HRM) mechanism was proposed to introduce HHC algorithm into FSL models.
- •
Based on HRM mechanism, a Hyperbolic Measurement Network (HM-Net) was designed for FSL task of traffic signs.Due to lighting intensity, shooting angle, and motion blur, many traffic sign images exhibit significant feature degradation, which greatly reduces the domain adaptation ability of deep learning models in practical scenarios. This phenomenon is more prominent in Few Shot Learning (FSL) with extremely limited data. To address this issue, this study proposes an improved FSL model. First, this study introduces hyperbolic space theory and designs a Hyperbolic Hierarchical Constraint (HHC) algorithm to acquire domain knowledge by leveraging neighborhood constraints in hyperbolic space. Then, a Hyperbolic Relationship Metric (HRM) mechanism based on the HHC algorithm is proposed to improve the model’s generalization ability across data samples. Finally, based on the HRM mechanism, an FSL model named Hyperbolic Measurement Network (HM-Net) is constructed for feature-degraded traffic sign images. For FSL tasks, HM-Net’s performance on the benchmark dataset exceeds that of existing models. For FSL-based traffic sign detection tasks, HM-Net achieves higher performance than existing models. In cross-domain experiments, HM-Net demonstrates stronger domain adaptation capability than existing models.
評(píng)論 0