GE-YOLOv8s: improved YOLOv8 lightweight air cargo inspection model

As deep learning technology advances, neural networks have increasingly been applied to aerospace target detection. In computer vision, the prevailing trend is toward deeper and more complex models to boost accuracy [1]. However, these improvements often compromise efficiency in terms of size and speed, particularly when constrained by computational resources. Meeting large-scale computation needs thus becomes a significant challenge. To address this, we propose GE-YOLOv8s, a lightweight YOLOv8 model tailored for air cargo detection. This design preserves detection accuracy while minimizing model parameters and computational demands, Rendering it appropriate for edge computing and real-time applications. The model uses a lightweight backbone network, GhostNet, To swap out the backbone of YOLOv8. To minimize the model parameters and computation even further, this study replaces some of the convolutional modules in YOLOv8 with GhostConv convolutional modules, and replaces the feature fusion module c2f with C3Ghost module. Additionally, the feature extraction ability of the model is improved by introducing the lightweight attention mechanism ECA module. The results of the experiment indicate that the GE-YOLOv8s model achieves 96% detection accuracy and 97% recall. In comparison to the original YOLOv8s network, the detection accuracy decreases by only 1.4% and the recall rate decreases by only 1.3%, while the amount of model parameters decreases by 46.5% and the amount of computation (measured in GFLOPs) decreases by 42.7%. The GE-YOLOv8s model greatly reduces parameters and computation while maintaining accuracy, fulfilling lightweight and real-time detection requirements.