STEdge: Self-Training Edge Detection With Multilayer Teaching and Regularization.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Supervised edge detection methods require extensive pixel-wise annotations, which are time-consuming and costly to acquire.
• Leveraging large-scale unlabeled image datasets offers a promising alternative to reduce annotation dependency.

Purpose of the Study:

• To develop a self-supervised framework for edge detection that mitigates the need for manual pixel-wise annotations.
• To improve the performance and cross-dataset generality of edge detection models.

Main Methods:

• A self-supervised framework incorporating multilayer regularization and self-teaching was designed.
• Consistency regularization was applied using L0-smoothing as a perturbation to enforce consistent outputs across image variations.
• Multilayer supervision was achieved using pseudo-labels, initially Canny edges, iteratively refined by the network.

Main Results:

• The proposed method achieved a favorable balance between precision and recall, outperforming traditional supervised approaches.
• Significant performance improvements were observed with minimal refinement on target datasets.
• The method demonstrated strong cross-dataset generality, enhancing existing edge detectors.

Conclusions:

• Self-supervised learning, through multilayer regularization and self-teaching, offers an effective approach to edge detection without extensive manual labeling.
• The framework improves edge detection accuracy and robustness, showing potential for broad application in computer vision tasks.