Enhancing Deep Edge Detection through Normalized Hadamard-Product Fusion.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Existing deep edge detection methods like Holistic Edge Detection (HED) combine multiple feature side outputs (SOs).
• These methods often neglect diverse edge importance within a single output, leading to increased noise or thick edges.
• This results in a trade-off between detecting desired edges and accepting unwanted noise.

Purpose of the Study:

• To introduce a new deep network approach for edge detection using a normalized Hadamard-product (NHP) operation.
• To address the limitations of existing methods in handling diverse edge importance and noise.
• To improve the accuracy and clarity of edge maps generated by deep learning models.

Main Methods:

• Proposing a novel normalized Hadamard-product (NHP) operation-based deep network for edge detection.
• Utilizing the Hadamard-product to multiply side outputs from the backbone network, promoting feature agreement across scales.
• Generating additional Mutually Agreed Salient Edge (MASE) maps to enhance hierarchical feature representation without added complexity.

Main Results:

• The NHP operation significantly improves edge detection performance.
• Achieved an Optimal Dataset Scale (ODS) score of 0.818 on the BSDS500 dataset.
• Outperformed human performance (0.803) and achieved state-of-the-art results in deep edge detection.

Conclusions:

• The NHP operation effectively suppresses weak, disagreed signals while encouraging agreement among features at different scales.
• This approach leads to cleaner and more accurate edge maps compared to existing methods.
• The proposed NHP-based network represents a significant advancement in deep edge detection technology.