BDCN: Bi-Directional Cascade Network for Perceptual Edge Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Edge detection is crucial for understanding image content at various scales.
• Existing methods struggle with effectively utilizing multi-scale representations for accurate edge detection.

Purpose of the Study:

• To develop a novel deep learning architecture for robust multi-scale edge detection.
• To improve the accuracy and efficiency of edge detection algorithms.

Main Methods:

• Proposed a bi-directional cascade network (BDCN) architecture.
• Introduced a scale enhancement module (SEM) using dilated convolutions for multi-scale feature enrichment.
• Implemented scale-dedicated layer supervision for improved representation learning.

Main Results:

• Achieved an ODS F-measure of 0.832 on the BSDS500 dataset, surpassing state-of-the-art by 2.7%.
• Demonstrated a compact network with fewer parameters due to scale-dedicated layers.
• Showcased performance improvements in downstream tasks like image segmentation and optical flow estimation.

Conclusions:

• The BDCN architecture effectively learns multi-scale representations for superior edge detection.
• The proposed methods enable accurate delineation of edges across different scales.
• The approach offers a computationally efficient and effective solution for edge detection and related computer vision tasks.