Lightweight convolutional neural network for UAV onboard visual anomaly detection and localization

Unmanned aerial vehicle platforms increasingly demand efficient visual anomaly detection capabilities for infrastructure inspection and environmental monitoring applications. However, deploying sophisticated deep learning models on resource-constrained embedded processors remains challenging due to computational limitations. This paper proposes a lightweight convolutional neural network architecture specifically designed for UAV onboard anomaly detection and localization. The framework centers on a Ghost-Separable Block that integrates ghost convolution, depthwise separable convolution, and coordinate attention into a single unit tailored for aerial imagery; rather than a wholly new operator, it is a deliberate combination whose value we establish empirically, trimming parameters by 26% while preserving feature discrimination. The embedded coordinate attention preserves directional spatial information essential for accurate target localization in aerial imagery. The bidirectional feature pyramid network enables effective multi-scale fusion to handle scale variations inherent in UAV-captured scenes. A dual-branch prediction head with Complete Intersection over Union loss performs simultaneous anomaly classification and spatial regression. Extensive experiments on a collected UAV aerial anomaly dataset demonstrate that the proposed method achieves 81.2% mean average precision at 0.5 IoU threshold with only 2.18 million parameters, surpassing mainstream lightweight detectors such as YOLOv8n and EfficientDet-D0: it raises mean average precision while cutting parameters by 31% and computation by 59% relative to YOLOv8n. The dual-branch head and the Complete Intersection over Union loss are adopted rather than newly proposed; the block arrangement is where the contribution lies. Beyond the self-collected data, the model preserves its accuracy-efficiency advantage on the public VisDrone benchmark, and on a Jetson Xavier NX it sustains 42 FPS natively and 67 FPS after TensorRT FP16 optimization within a 15 W power envelope, confirming real-time onboard feasibility.