Foreground Capture Feature Pyramid Network-Oriented Object Detection in Complex Backgrounds.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Feature pyramids are crucial for multiscale object detection.
• Existing methods struggle with complex backgrounds, leading to suboptimal foreground feature capture.
• This limits the discriminative power of multiscale semantic features.

Purpose of the Study:

• To propose a Foreground Capture Feature Pyramid Network (FCFPN) for robust multiscale object detection.
• To address the challenge of inadequate feature learning in complex background environments.
• To enhance the capture of discriminative multiscale foreground semantic features.

Main Methods:

• The proposed FCFPN integrates a Foreground Dual Attention (FDA) module and a Pathway Aggregation (PA) structure.
• FDA enhances foreground channel and spatial features by activating responses and location features.
• PA adaptively learns fusion weights for multiscale features, improving semantic complementarity.

Main Results:

• FCFPN demonstrated superior detection average precision (AP) and feature learning performance.
• Evaluations on public and custom datasets confirmed the method's effectiveness.
• The approach successfully retained relevant feature information and suppressed conflicting data.

Conclusions:

• The Foreground Capture Feature Pyramid Network (FCFPN) effectively overcomes background interference in object detection.
• The FDA and PA modules significantly improve the learning of multiscale foreground features.
• FCFPN offers a promising solution for enhanced object detection in challenging visual scenes.