Scale Enhancement Pyramid Network for Small Object Detection from UAV Images.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Object detection in large-scale Unmanned Aerial Vehicle (UAV) imagery presents challenges, particularly for small objects with scale variations.
• Existing multi-scale feature fusion methods in feature pyramid networks (FPNs) struggle to capture long-range dependencies crucial for detecting objects of varying sizes.

Purpose of the Study:

• To develop a novel network, the Scale Enhancement Pyramid Network (SEPNet), to address the limitations of current object detection methods for UAV imagery.
• To improve the detection of small objects with significant scale variation by enhancing feature representation and preserving spatial information.

Main Methods:

• Proposed the Scale Enhancement Pyramid Network (SEPNet), incorporating a Context Enhancement Module (CEM) and a Feature Alignment Module (FAM).
• The CEM utilizes multi-scale atrous and multi-branch grouped convolutions to model global relationships and enrich feature representation.
• The FAM adaptively learns pixel offsets to maintain feature consistency and mitigate information conflicts during feature fusion.

Main Results:

• SEPNet achieved a 7.1% higher Average Precision (AP) score on the VisDrone dataset compared to state-of-the-art detectors, reaching 18.9% AP.
• The network demonstrated effectiveness in handling small objects with significant scale variations in complex UAV scenes.
• Achieved 81.5% AP on the PASCAL VOC dataset, indicating strong generalizability.

Conclusions:

• SEPNet offers a simple yet effective solution for object detection in challenging UAV imagery, outperforming existing methods.
• The proposed CEM and FAM modules successfully enhance feature representation and preserve spatial information, crucial for detecting small, varied-scale objects.