A Domain Adaptation Method for Object Detection in UAV Based on Semi-Supervised Learning

In recent years, the objects detection algorithm of deep learning neural network has been widely used in scene monitoring fields such as intelligent transportation and ground observation. But scene changes always happen caused by the seasons and the ground feature cover change, which makes the training sample (source domain) different from the application scene (target domain), and causes the original effective model to generate a lot of false alarms. This paper proposes a domain adaptation algorithm for Unmanned Aerial Vehicle (UAV) objects detection based on semi-supervised learning. Firstly applying the source domain model on target domain to generate pseudo-labels for semi-supervised learning, during fine-tuning the model with Pseudo-Label, the scale-aware loss is introduced to suppress false alarms in abnormal scale, which realize the domain adaptability of network model migration from source domain to target domain. Experiments on our dataset for monitoring ground objects show that: the proposed method avoids the catastrophic forgetting caused by the direct domain adaptation method without denoising, and maintains the target detection rate of the network model in different scenarios.