Edge Detection via Fusion Difference Convolution

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Edge detection is fundamental to computer vision.
• Current deep convolutional neural network (CNN) methods often rely on pre-trained networks and include background noise.
• A need exists for more robust and accurate edge detection models.

Purpose of the Study:

• To develop a novel edge detection model that overcomes limitations of existing CNN-based approaches.
• To improve the recognition of semantic and edge information in images.
• To create a model that can be trained from scratch and generalizes well.

Main Methods:

• Proposed four new fusion difference convolution (FDC) structures integrating traditional gradient operators into CNNs.
• Incorporated a channel spatial attention module (CSAM) and an up-sampling module (US).
• Trained the model from scratch on the BIPED dataset without pre-trained weights.

Main Results:

• Achieved promising results on the BIPED dataset.
• Demonstrated effective recognition of semantic and edge information.
• Showcased strong generalization capabilities to other datasets without fine-tuning.

Conclusions:

• The proposed FDC structures, CSAM, and US modules enhance edge detection performance.
• Training from scratch is viable and yields competitive results.
• The model offers a robust and generalizable solution for edge detection tasks.