<p>Automatic detection of foreign objects on power lines is crucial for preventing power outages, yet faces challenges posed by irregular shapes, severe occlusions, and category imbalances. Existing lightweight detectors often struggle with accuracy, robustness, or efficiency. To address this issue, we propose a lightweight, efficient object detection algorithm: LSP-YOLO. This method introduces three improvements over YOLOv10: (1) It incorporates a deformable convolution (LDConv) to enable adaptive feature extraction for irregularly shaped objects through a flexible sampling mechanism; (2) it designs an improved ASDI feature fusion module, using weighted Hadamard products and scaling factors to enhance the stability and robustness of feature interactions; (3) it proposes a novel loss function, IPIoU, which further introduces boundary direction error modeling on top of Inner-IoU, effectively improving localization accuracy in occlusion scenarios. To address the class imbalance issue, this paper constructs and expands a private dataset containing four types of foreign objects (bird nests, balloons, kites, and rubbish). Experimental results show that LSP-YOLO achieves 93.4 percent precision, 82.2 percent recall, and 88.1 percent mAP@50 under conditions of only 1.7 million parameters and 6.7 GFLOPs, significantly outperforming similar lightweight models such as YOLOv10n and YOLOv12n. Compared to mainstream detection methods, LSP-YOLO strikes a balance between lightweight design and high detection accuracy and robustness, demonstrating strong practical application value.</p>