DPED: Bio-inspired dual-pathway network for edge detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Edge detection is crucial for high-level visual tasks.
• Convolutional Neural Networks (CNNs) dominate current encoder-decoder edge detection methods.
• Swin Transformer (Swin) shows promise as an alternative to CNNs in visual tasks.

Purpose of the Study:

• To design a novel two-pathway encoding network for edge detection.
• To leverage Swin Transformer and biological vision principles.
• To improve performance and efficiency in edge detection.

Main Methods:

• A two-pathway encoding network combining a fine-tuned Swin Transformer and deep separable convolution.
• A second-pathway attention module for inter-pathway communication.
• A pathways fusion module for feature integration.

Main Results:

• The proposed method surpasses the CNN-based state-of-the-art (SOTA) method BDCN on the BSDS500 dataset.
• It demonstrates competitive performance against the Transformer-based SOTA method EDTER.
• The method offers advantages in terms of Floating Point Operations (FLOPs) and Frames Per Second (FPS) compared to EDTER.

Conclusions:

• The novel two-pathway network effectively enhances edge detection capabilities.
• The integration of Swin Transformer and CNN components provides a powerful hybrid approach.
• The proposed method offers a more efficient and performant solution for edge detection tasks.