An advanced adaptive detector for oriented objects in remote sensing imagery.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Object detection models like YOLOv8 struggle with oriented objects in remote sensing due to scale variations, irregular shapes, and arbitrary orientations.
• Existing methods often fail to address multiple challenges simultaneously, limiting oriented detection performance.
• Inconsistency in loss measurement and angle discontinuity are significant hurdles in current oriented detection techniques.

Purpose of the Study:

• To develop an advanced adaptive detector, AdaR-YOLOv8, enhancing YOLOv8 for oriented object detection in remote sensing imagery.
• To overcome limitations of existing methods by proposing novel strategies for feature extraction, fusion, and loss calculation.
• To establish a new benchmark for oriented object detection accuracy in challenging remote sensing datasets.

Main Methods:

• Introduced the C2f-DCN module using deformable convolutions for dynamic shape adaptation during feature extraction.
• Implemented Multi-Scale Sequential Feature Fusion (MSFF) with 3D convolution for improved multi-scale feature merging and small target detection.
• Incorporated the KFIoU loss function to effectively handle angle regression issues and angle discontinuity.

Main Results:

• AdaR-YOLOv8 demonstrated significant improvements in detection accuracy on the DOTA and HRSC2016 datasets.
• The proposed methods successfully addressed challenges related to scale variation, irregular shapes, and arbitrary orientations.
• Achieved state-of-the-art performance, establishing a new benchmark for oriented object detection in remote sensing.

Conclusions:

• AdaR-YOLOv8 offers a robust solution for oriented object detection in remote sensing imagery.
• The integration of deformable convolutions, advanced feature fusion, and specialized loss functions enhances detection capabilities.
• This work advances the field of remote sensing image analysis by providing a more accurate and reliable detection system.