Transmission Line Foreign Object Detection (TLFOD) faces three key challenges. Traditional attention mechanisms are inefficient in feature selection within complex backgrounds. Standard pooling operations result in the loss of high-frequency detail information. Traditional feature pyramids struggle to adapt to multi-scale targets. To address these challenges, this paper proposes the YOLO-TL model based on YOLOv11n. Firstly, the Convolutional Pooling Gated Linear Unit (CPGLU) is designed to replace the traditional C3K2 structure. The module effectively suppresses the background noise of foreign objects on transmission lines. Secondly, the WaveletPool algorithm is proposed. It maintains feature integrity in the frequency domain, preserving critical details of small objects and suppressing high-frequency noise by selectively discarding high-frequency subbands. Thirdly, the Lightweight Multi-Scale Adaptive Feature Pyramid Network (LMAFPN) is designed to enhance multi-scale object detection performance through weighted feature fusion, efficient upsampling convolution blocks, and multi-scale convolution modules. Finally, the Task Synchronization Head (TSHead) is designed to promote the collaboration of classification and localization tasks through group-normalized sharing mechanisms and task decomposition. Experimental results demonstrate improvements over the original YOLOv11n on the RailFOD23 dataset, with mAP \(_{50}\) achieving 95.0%. Parameters are reduced by 46.5%. FLOPs are reduced by 26.9%. Size is reduced by 41.5%. In addition, the YOLO-TL model also shows strong generalization ability on the SPGFOD dataset. Due to its high detection performance and lightweight design, the model provides a highly cost-effective solution for edge-side power inspection applications.