Overcoming Scale Variations and Occlusions in Aerial Detection: A Context-Aware DEIM Framework.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Remote Sensing

Background:

• Object detection in Unmanned Aerial Vehicle (UAV) imagery is crucial for applications like railway inspection and waste management.
• Existing detectors like DEIM face challenges with weak feature responses and spatial misalignment in aerial data.

Purpose of the Study:

• To propose SCA-DEIM, a context-aware, real-time detection framework for UAV imagery.
• To enhance the detection of small objects and improve feature extraction and spatial alignment.

Main Methods:

• Introduced the Adaptive Spatial and Channel Synergistic Attention (ASCSA) module to amplify faint small-target signals.
• Developed the Cross-Stage Partial Shifted Pinwheel Mixed Convolution (CSP-SPMConv) to align receptive fields and fuse features across scales.
• Utilized the VisDrone2019, UAVVaste, and UAVDT datasets for comprehensive evaluation.

Main Results:

• SCA-DEIM achieved a 1.8% increase in Average Precision (AP), 2.3% in AP for small objects (APs), and 2.0% in AP for large objects (APl) on VisDrone2019.
• The model demonstrated competitive inference speed and strong robustness under varying illumination conditions.
• Further validation confirmed enhanced small object detection performance on UAVVaste and UAVDT datasets.

Conclusions:

• SCA-DEIM effectively addresses challenges in UAV object detection, particularly for small objects.
• The proposed ASCSA and CSP-SPMConv modules significantly improve feature extraction and spatial alignment.
• The framework offers a robust and efficient solution for real-time aerial object detection tasks.