In response to the issues of low accuracy and poor real-time performance caused by the large scale of foreign objects and complex backgrounds in existing transmission line foreign object detection algorithms, a lightweight model based on improved YOLOv8 is proposed. Firstly, a modified feature extraction network is designed by integrating a dual-branch architecture with attention mechanisms into MobileNetV2, optimizing the model's parameters and enhancing the feature representation capability. Additionally, the CBAM attention mechanism is embedded into the SPPF and PANet modules to enhance the discrimination between foreign objects and the environment, to improve the model's detection capability in complex backgrounds and consequently enhance the detection accuracy. Furthermore, the WIoU loss function is introduced to properly allocate gradient gains, enhancing the model's generalization ability and detection box localization accuracy. Experimental results demonstrate that the improved YOLOv8 achieves a detection accuracy of 96.12% mAP and an inference speed of 60 frames per second. It outperforms YOLOv8 and five other mainstream object detection models in terms of overall performance. Moreover, it exhibits high stability and robustness under different lighting conditions, proving its practicality in complex environments. Further testing on embedded devices confirms that even under limited computational resources, the improved YOLOv8 can still achieve accurate detection.