A Lightweight Transformer-Based FPN for Efficient SMA RF Connector Insertion Quality Monitoring

Abstract In SMA RF connector insertion quality inspection, the coexistence of small target features, complex backgrounds, and strict real-time requirements poses significant challenges for existing detection frameworks. Traditional Feature Pyramid Networks (FPN) effectively fuse multi-scale features but rely solely on convolutional operations, limiting their ability to model global dependencies and reducing robustness in complex industrial scenes. To address these issues, we propose a novel framework that integrates a lightweight Transformer encoder into the FPN. The encoder replaces the traditional multi-head self-attention with a multi-head hybrid sparse attention mechanism, reducing computational complexity by 26.4% while enhancing global–local feature modeling. In addition, a dynamic feed-forward network with grouped convolution improves local feature representation and reduces parameters by 34.7%. Experimental results on the SMA RF connector dataset demonstrate significant performance gains: detection accuracy for loose, tight, and empty connections improves by 40.8%, 38.1%, and 16.3%, respectively, and inference speed increases by 20.5%. To verify engineering feasibility, we developed a complete graphical user interface (GUI) system integrating image acquisition, detection inference, and result visualization. The proposed framework delivers an effective balance between accuracy and efficiency, showing strong potential for deployment in real-world industrial quality inspection.