Elevating Detection Performance in Optical Remote Sensing Image Object Detection: A Dual Strategy with Spatially...

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Object detection in optical remote sensing images faces challenges with discontinuous boundaries, limiting accuracy at high Intersection over Union (IoU) thresholds.
• Existing methods struggle to precisely optimize the angle parameters of rotated bounding boxes, hindering performance in complex scenarios.

Purpose of the Study:

• To introduce a novel network, the Spatial Adaptive Angle-Aware (SA3) Network, designed to enhance object detection accuracy in optical remote sensing.
• To address the limitations of discontinuous boundaries and improve performance specifically at high IoU thresholds.
• To develop an effective angle regression loss function for rotated bounding boxes.

Main Methods:

• Proposed the Spatial Adaptive Angle-Aware (SA3) Network with a hierarchical refinement approach (coarse regression, fine regression, precise tuning) for rotated bounding box angle optimization.
• Introduced a Gaussian transform-based IoU factor and developed the Edge-aware Skewed Bounding Box Loss (EAS Loss) to enhance angle regression.
• Implemented class-aware and class-agnostic strategies for task-specific adaptation.

Main Results:

• The SA3 Network significantly improved detection accuracy, particularly at high IoU thresholds.
• The EAS Loss enhanced the loss gradient in the final stage of angle regression, increasing training efficiency and metric alignment.
• The combined SA3 Network and EAS Loss elevated the mean Average Precision (mAP) of the ReBiDet model on DOTA-v1.5 to 78.85%, with substantial gains under high IoU conditions.

Conclusions:

• The SA3 Network effectively addresses challenges in rotated object detection by refining angle parameters.
• The EAS Loss provides a significant improvement in training efficiency and accuracy for angle regression.
• The proposed methods substantially enhance the performance of existing object detection models like ReDet and ReBiDet, especially in demanding high IoU scenarios.