GS-YOLO: A Lightweight Identification Model for Precision Parts

With the development of aerospace technology, the variety and complexity of spacecraft components have increased. Traditional manual and machine learning-based detection methods struggle to accurately and quickly identify these parts. Deep learning-based object detection networks require significant computational resources and high hardware requirements. This study introduces Ghost SCYLLA Intersection over Union You Only Look Once (GS-YOLO), an improved image recognition model derived from YOLOv5s, which integrates the global attention mechanism (GAM) with the Ghost module. The lightweight Ghost module substitutes the original convolutional layers, producing half of the features via convolution and the other half by symmetric linear operations. This minimizes the computing burden and model parameters by effectively acquiring superfluous feature layers. A more lightweight SimSPPF structure is created to supplant the old spatial pyramid pooling—fast (SPPF), enhancing the network speed. The GAM is included in the bottleneck architecture, improving feature extraction via channel–space interaction. The experimental results on the custom-made precision component dataset show that GS-YOLO achieves an accuracy of 96.5% with a model size of 10.8 MB. Compared to YOLOv5s, GS-YOLO improves accuracy by 1%, reduces parameters by 23%, and decreases computational requirements by 40.6%. Despite the model’s light weight, its detection accuracy has been improved.