Distance Field-Based Convolutional Neural Network for Edge Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Convolutional Neural Networks (CNNs) are effective in image processing.
• Edge detection is a fundamental computer vision task.
• Improving edge detection accuracy is an ongoing research challenge.

Purpose of the Study:

• To propose an accurate edge detector using a novel distance field-based convolutional neural network (DF-CNN).
• To enhance end-to-end object edge detection accuracy through deep learning.
• To validate the effectiveness of the proposed distance field branch.

Main Methods:

• Developed a DF-CNN integrating a feature extraction backbone and a supervised distance field branch.
• The distance field branch predicts Euclidean distance from non-edge to nearest edge points.
• Network trained by minimizing a weighted sum of distance field loss and cross-entropy loss.

Main Results:

• The proposed DF-CNN achieves superior performance compared to existing edge detection methods.
• Experimental results demonstrate the effectiveness of the distance field branch in improving accuracy.
• The network successfully performs accurate end-to-end object edge detection.

Conclusions:

• The proposed DF-CNN is an effective approach for accurate object edge detection.
• The distance field branch significantly contributes to improved edge detection accuracy.
• This deep learning framework offers a promising direction for future edge detection research.