Multi-Scale pattern-Aware task-Gating network for aerial small object detection.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Remote sensing images (RSIs) present unique challenges for object detection, including complex backgrounds, cluttered object distributions, small object sizes, and scale variations.
• Existing methods struggle to effectively address these challenges due to limitations in capturing multi-scale spatial patterns and fusing semantic information.

Purpose of the Study:

• To develop an advanced object detection network specifically designed for the complexities of remote sensing images.
• To improve the accuracy and robustness of object detection in challenging remote sensing scenarios.

Main Methods:

• Proposed a Multi-Scale Pattern-Aware Network (MPNet) backbone using complementary small and large kernel convolutions for comprehensive feature extraction.
• Introduced a Multi-Head Cross-Space Encoder (MCE) to enhance semantic fusion and pixel-level spatial sensitivity using directional strip convolutions.
• Developed a Dynamic Task-Gating (DTG) head to adaptively recalibrate feature representations between classification and localization tasks.

Main Results:

• The proposed Multi-Scale Pattern-Aware Task-Gating Network (MPTNet) achieved high performance on VisDrone, DIOR, and COCO-mini datasets.
• Achieved AP50 scores of 43.3% on VisDrone, 80.6% on DIOR, and 49.5% on COCO-mini.
• Demonstrated significant improvements in object detection accuracy for remote sensing applications.

Conclusions:

• MPTNet effectively addresses the challenges of object detection in remote sensing images.
• The combination of MPNet, MCE, and DTG contributes to superior performance in complex scenes.
• The proposed method offers a promising solution for various military and civilian applications relying on remote sensing object detection.