CycleGAN-Assisted Domain Adaptation for UAV Payload Detection

As uncrewed aerial vehicles (UAVs) become increasingly integrated into various domains, concerns about their potential misuse have grown, highlighting topics such as reliable detection of carrying payloads. This paper presents a novel two-stage deep learning framework for detecting loaded UAVs using vision-based classification. Due to the high cost of collecting real-world training data, we utilize a synthetic dataset for model training. However, domain shifts between synthetic and real-world data can degrade classification performance. To address this challenge, we employ a CycleGAN-based domain adaptation method that transforms real test samples into their synthetic counterparts, ensuring consistency with the training distribution. The adapted samples are then classified using a pre-trained deep network based on ResNet and EfficientNet architectures. By preserving the integrity of the classifier while bridging the domain gap, our approach significantly improves UAV payload detection. Experimental results demonstrate the effectiveness of the proposed method in enhancing classification accuracy for real-world UAV monitoring applications.