UAVDet: A Lightweight Fast Detection Model for Marine Ships based on Edge Mobile Devices

Abstract

Recently, Unmanned Aerial Vehicles have been widely used in the fields of water traffic supervision and maritime sovereignty inspection, becoming an important means of data acquisition. It is crucial to apply deep learning-based target detection technology to UAV edge devices. Traditional detectors are often underperformed when deployed directly to UAVs. One reason is that the amount of UAV imagery data is often limited and insufficient to support the training of deep learning algorithms. The second is that deep learning-based detectors often have huge models and huge amounts of parameters with high computational complexity, making it difficult to deploy them to work effectively on edge mobile devices with extremely limited computational resources and memory. To solve these problems, we proposed a new detection model for UAV view ship target based on YOLOv4. To this end, first, we constructed a satellite remote sensing ship image dataset and used transfer learning to reduce the reliance on model training data. Second, we lightened the model by sparsity training, channel and layer pruning, and then used knowledge distillation techniques to rebound the accuracy. In the end, the model size is reduced by 97.19% and the detection time of a single image is reduced by 39.73% while maintaining high detection accuracy, achieving high precision real-time detection suitable for deployment on edge devices such as UAVs.