Gelan-SE: Squeeze and Stimulus Attention Based Target Detection Network for Gelan Architecture

In the field of computer vision, the development of efficient target detection models for resource-constrained environments represents a significant challenge. The objective of this study is to optimize and enhance the Generalized Efficient Layer Aggregation Network (GELAN) architecture of YOLOv9, thereby improving the model’s performance and efficiency. Although the GELAN architecture of YOLOv9 has already demonstrated robust performance, this study achieves further optimization through the development of the RepNCSPELAN4WithSE module and the Gelan-SE-Head network structure. The study preserves the RepNCSPELAN4 module of the original GELAN architecture while integrating Deep Separable Convolution (DWConv). Moreover, to further enhance the model performance, this study introduces the SE (Squeeze-and-Excitation) attention mechanism and the SAM (Spatial Attention Module) attention mechanism in the backbone or head network. The experimental results demonstrate that the Gelan-SE-Head network, in comparison to the original GELAN architecture, exhibits a reduction in the number of model parameters by 9.30%, an enhancement in the inference speed by 6.29%, and an improvement in the mAP50 and mAP50-95 metrics by 0.62% and 0.91%, respectively. These enhancements demonstrate that the experimentally developed attention mechanism and network architecture can not only enhance the accuracy of the target detection model, but also markedly reduce the number of model parameters and improve computational efficiency in a constrained environment.