PSMDet: Enhancing Detection Accuracy in Remote Sensing Images Through Self-Modulation and Gaussian-Based Regression.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Conventional object detection struggles with multi-scale, high aspect ratio, and arbitrarily oriented targets in optical remote sensing images (ORSIs).
• Existing methods face challenges in feature extraction and bounding box regression for complex ORSI targets.

Purpose of the Study:

• To propose a novel detection framework, the Progressive Self-Modulating Detector (PSMDet), to address limitations in ORSI object detection.
• To enhance feature extraction, alignment, and bounding box regression for complex targets in ORSIs.

Main Methods:

• Developed PSMDet incorporating self-modulation at backbone, feature pyramid network (FPN), and detection head stages.
• Utilized a reparameterized large kernel network (RLK-Net) for enhanced multi-scale feature extraction.
• Introduced an adaptive perception network (APN) with self-attention for feature alignment, a Gaussian-based bounding box representation, and smooth relative entropy (smoothRE) loss for regression.

Main Results:

• PSMDet achieved high performance on HRSC2016 and UCAS-AOD datasets, with mean Average Precision (mAP) scores of 90.69% and 89.86%, respectively.
• The framework demonstrated robust performance in detecting complex targets in ORSIs.

Conclusions:

• PSMDet offers a significant advancement in object detection for ORSIs, addressing key challenges.
• The proposed framework is adaptable for various applications requiring high-precision object detection, including autonomous driving and industrial defect detection.