SemiContour: A Semi-supervised Learning Approach for Contour Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Supervised contour detection demands extensive labeled training data, which is often impractical or costly.
• Existing methods face limitations due to the scarcity of annotated datasets.

Purpose of the Study:

• To develop a semi-supervised learning (SSL) approach for contour detection using limited labeled training data.
• To enhance the performance of structured random forests (SRF) by incorporating unlabeled data.

Main Methods:

• Proposed a semi-supervised structured ensemble learning framework based on structured random forests (SRF).
• Introduced a sparse representation technique for unsupervised learning of image patch structures.
• Developed a novel and efficient sparse coding algorithm to improve learning speed.

Main Results:

• Achieved competitive contour detection accuracy with only three labeled images.
• Demonstrated the effectiveness of the proposed SSL method on the BSDS500 and NYU Depth datasets.
• Validated the superiority of the approach compared to traditional methods.

Conclusions:

• Semi-supervised learning is a viable and effective strategy for contour detection with limited data.
• The proposed sparse representation and novel sparse coding algorithm significantly enhance learning efficiency and accuracy.
• This work pioneers the application of SSL in the field of contour detection.