DCA-Net: Advanced Small Object Detection Scheme for UAVs Using Contextual Feature Extraction

Detecting small-scale infrared objects such as UAVs in complex scenarios remains a significant challenge due to low feature resolution and limited contextual information, particularly when dealing with drones and other flying objects. To address this issue, we have proposed a Dynamic Context Attention Network (DCA-Net) model, designed to enhance detection accuracy and computational efficiency. DCA-Net integrates Depthwise Convolutions (DWConv) for reduced computational complexity, Efficient Channel Attention (ECA) and C3-ECA modules for dynamic channel attention to effectively capture complex features, and Efficient Multi-Receptive Pooling (EMRP) layers for processing diverse spatial hierarchies. This novel approach improves both feature representation and processing efficiency. Experimental evaluations on a multi-class infrared dataset demonstrate that DCA-Net outperforms the original YOLOv5 model, achieving 2.1% higher mean Average Precision (mAP), while reducing the complexity by 29% in terms of GFLOPs. The results highlight the effectiveness of DCA-Net in overcoming the limitations of existing object detection models, offering superior performance and real-time inference capabilities.