MSB-FCN: Multi-Scale Bidirectional FCN for Object Skeleton Extraction

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) have advanced object skeleton detection (OSD).
• Current CNN-based OSD methods often use skip-layer structures, combining shallow and deep features.
• Shallow features can be noisy and lack semantic information, leading to OSD inaccuracies.

Purpose of the Study:

• To propose a novel network architecture, the Multi-Scale Bidirectional Fully Convolutional Network (MSB-FCN), for improved OSD accuracy.
• To enhance the gathering and utilization of multi-scale high-level contextual information.
• To overcome the limitations of noisy shallow features in existing OSD methods.

Main Methods:

• Developed the Multi-Scale Bidirectional Fully Convolutional Network (MSB-FCN).
• Utilized only deep features for multi-scale representation construction.
• Incorporated a bidirectional structure for enhanced contextual knowledge capture.
• Integrated dense connections and an attention pyramid to refine feature learning and propagation.

Main Results:

• The MSB-FCN effectively gathers and enhances multi-scale high-level contextual information.
• The network learns semantic-level information from different sub-regions.
• Dense connections ensure cross-scale information encoding, and the attention pyramid reduces unreliable features.
• Achieved significant improvements over state-of-the-art OSD algorithms on various benchmarks.

Conclusions:

• The proposed MSB-FCN architecture significantly enhances object skeleton detection accuracy.
• Relying solely on deep features with a bidirectional, multi-scale approach is effective.
• The integration of dense connections and attention mechanisms further boosts performance by refining feature propagation and reducing noise.