A CNN-transformer dual-branch network with structure-aware loss for high-resolution edge detection

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Edge detection is crucial for computer vision tasks like segmentation and object detection.
• Current deep learning methods use pixel-wise losses (e.g., Binary Cross Entropy) that ignore edge geometry.
• This pixel-level supervision leads to artifacts like fragmented or blurred edges.

Purpose of the Study:

• To develop a novel loss function that explicitly supervises edge geometric properties.
• To improve the structural coherence and accuracy of deep learning-based edge detection.
• To address the limitations of existing pixel-wise loss functions in edge detection.

Main Methods:

• Proposed a novel Edge-Structure-Aware Loss Function incorporating gradient, continuity, and directional consistency constraints.
• Introduced a Three-Stage Dynamic Loss Scheduling Strategy using curriculum learning.
• Designed a dual-branch network architecture integrating semantic and contextual features.

Main Results:

• Achieved competitive performance on multiple benchmarks, including BSDS500 (ODS=0.847/OIS=0.861), Multicue (ODS=0.899/OIS=0.907), and NYUDv2 (ODS=0.761/OIS=0.776).
• Demonstrated significant improvements in edge map quality by addressing fragmentation and coherence issues.
• Validated the effectiveness of the proposed structure-aware loss and scheduling strategy.

Conclusions:

• The proposed Edge-Structure-Aware Loss Function effectively enhances edge detection accuracy and structural integrity.
• The dynamic loss scheduling and dual-branch architecture provide a robust framework for structure-aware edge learning.
• This work offers a significant advancement over traditional pixel-wise loss functions in computer vision.