Parallel Residual Bi-Fusion Feature Pyramid Network for Accurate Single-Shot Object Detection

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Feature Pyramid Networks (FP) are crucial for visual detection but struggle with localization accuracy due to pooling shifts.
• The effectiveness of FP diminishes with deeper backbones and hinders simultaneous detection of small and large objects.

Purpose of the Study:

• To propose a novel Parallel Residual Bi-Fusion Feature Pyramid Network (PRB-FPN) for fast and accurate single-shot object detection.
• To enhance feature retention and localization accuracy by introducing bi-directional fusion and improved feature processing.

Main Methods:

• Developed a parallel Feature Pyramid structure with bi-directional (top-down and bottom-up) fusion.
• Introduced a bottom-up fusion module (BFM) and a concatenation and re-organization (CORE) module for improved feature fusion and information recovery.
• Incorporated a residual design into the CORE module (Re-CORE) for enhanced training and backbone integration.

Main Results:

• The PRB-FPN effectively detects both small and large objects simultaneously with high accuracy.
• Bi-directional fusion recovers lost information from lower-layer feature maps, improving localization.
• The Re-CORE module facilitates integration with various deep or light backbones.

Conclusions:

• The proposed PRB-FPN significantly improves object detection performance, achieving state-of-the-art results on benchmark datasets.
• The network offers a robust solution for accurate and efficient single-shot object detection.
• The design advancements address key limitations of traditional Feature Pyramid Networks.