MSFC-Net: Multi-scale feature and contextual fusion network for high-precision UAV detection

Detecting Unmanned Aerial Vehicles (UAVs) in complex environments remains challenging due to scale variation and target occlusion. To enhance detection performance, this study proposes the Multi-Scale Feature and Contextual Fusion Network (MSFC-Net). To optimize the backbone, the Efficient Feature Fusion Convolution (EFFConv) module combines multi-scale convolutions with channel grouping. This approach enables simultaneous capture of fine-grained local details and global contextual information while maintaining computational efficiency. Building upon this foundation, the Contextual-information Fusion Pyramid Network (CIFPN) addresses existing feature fusion limitations. The Dual Conv Fusion (DCF) Block implements attention-guided feature fusion to optimally integrate shallow and deep layer features for enhanced object detection. The Content-Guided Attention (CGA) fusion module employs triple-attention mechanisms to dynamically weight multi-level contextual features. To further improve detection accuracy, the Multi-Scale Quality-Fused Detection Head (MSQF-Head) introduces a quality-aware detection paradigm. It incorporates shared convolutional operations for efficient multi-scale feature processing and a Location Quality Estimation (LQE) module. The LQE module dynamically adjusts confidence scores based on statistical corner probability analysis to improve bounding box localization accuracy. Comprehensive experiments across visible-spectrum DUT-Anti-UAV, constructed Bird-UAV, and infrared datasets demonstrate superior performance with validation on aerial imagery scenarios. Real-time deployment achieves 13 FPS on Jetson Orin Nano, proving applicability.