Partial feature reparameterization and shallow-level interaction for remote sensing object detection.

Area of Science:

• Computer Vision
• Deep Learning
• Remote Sensing

Background:

• Object detection in remote sensing images presents challenges in balancing performance and computational efficiency.
• Existing deep learning models often struggle to meet the demands of real-time processing and high accuracy.

Purpose of the Study:

• To develop an efficient one-stage object detector for remote sensing image analysis.
• To improve detection performance and computational efficiency in remote sensing object detection tasks.

Main Methods:

• Proposed PRepConvBlock for reduced complexity and expanded receptive fields using reparameterization convolution.
• Introduced SB-FPN, a shallow multi-scale fusion framework enhancing feature representation through cross-scale interaction.
• Developed the Shallow-level Optimized Reparameterization Architecture Detector (SORA-DET) for UAV applications.

Main Results:

• SORA-DET achieved 39.3% mAP50 on the VisDrone2019 test set and 84.0% mAP50 on the SeaDroneSeeV2 validation set.
• The detector boasts significantly fewer parameters (smaller than 88.1% of large models) and a fast inference speed of 5.4 ms.
• Outperformed many large-size models and state-of-the-art works in remote sensing object detection.

Conclusions:

• SORA-DET offers a compelling solution for efficient and accurate object detection in remote sensing imagery.
• The proposed innovations effectively address the trade-off between performance and computational cost.
• SORA-DET demonstrates strong potential for real-world UAV remote sensing applications.