Improved Lightweight YOLOv5 Using Attention Mechanism for Satellite Components Recognition

Recognizing typical components of the satellite is a challenging task for on-orbit services. This article proposes a YOLOv5-based satellite components recognition model (YSCRM) on a computationally limited platform. In particular, feature fusion layers and selective kernel networks (SK-Nets) are introduced to handle the complex multimodal recognition problem of the components, improving the model’s feature selection and representation capability and significantly increasing the recognition accuracy. The transformer encoder modules are used at the end of the Yolov5 Neck to explore the prediction potential with the self-attention mechanism. The channel sparseness training method is implemented to reduce the model size to deploy the model on an embedded platform. Moreover, for the purpose of enlarging the training datasets and their diversity, a data generation approach based on the synthetic images generated by 3ds Max and a data augmentation method based on cycle-consistent adversarial network (Cycle-GAN) are presented in this research. The proposed hybrid dataset is used for training and testing, and extensive qualitative and quantitative experiments are performed. The results demonstrate that the YSCRM has significant recognition performance, and the five typical components, solar panel, body, tripod, nozzle, and docking ring, can be recognized with high accuracy, which indicates that the proposed model can be used for real-world scenarios.