Infrared Drone Target Recognition Based on Global-Local Context Aggregation with Self-Attention

As drone usage increases, the security risks linked to unauthorized flights have become more evident. Infrared imaging detection technology, recognized for its robust anti-interference features and ability to provide continuous surveillance, effectively monitors drone activities in urban areas. To tackle the issues arising from the limited infrared characteristics of drones and their close resemblance to background elements-which leads to low real-time detection and recognition rates-we created a detailed urban infrared drone target dataset and introduced the Infrared Unmanned Aerial Vehicle You Only Look Once(IUAV-YOLO) network model. This model is based on global-local context aggregation with self-attention. Our method starts with a Backbone Feature Extraction Framework (BFEM) aimed at maintaining semantic information for small targets within deep networks, while a multi-branch convolution broadens the receptive field to improve local perception. We also present a self-attention context aware module Spatial Pyramid Pooling Fast Global Average Max(SPPF-GAM) to combine multi-scale contextual features, enhancing spatial awareness. Finally, the Spatial Context Aware Module(SCAM) and CSP Bottleneck With Two Convolutions(C2f) modules are linked to merge global and local context across both channel and spatial dimensions, effectively addressing the subtle features of small drones against intricate backgrounds. Experimental findings show that the IUAV-YOLO model improves the Mean Average Precision ([email protected]) by 4.3% compared to the baseline model, while also decreasing the number of parameters by 433,636. Overall, our proposed network demonstrates significant benefits over other leading real-time detection algorithms, making it an asset for enhancing security in urban settings.