Abstract: Power line inspection is a core part of power system operation and maintenance. The integration of drone technology and object detection algorithms has significantly improved inspection efficiency and safety, and is widely applied in China's power systems. However, detecting defects in tiny grid components (such as small fittings) in high-resolution drone images faces challenges such as tiny object scale and sparse spatial distribution. Direct application of detection models on full images leads to insufficient recall rates and low inference efficiency, failing to meet practical inspection needs. To address this challenge, a defect detection method for tiny power grid components based on multi-modal large model semantic guidance is proposed in this paper. The method first constructs a potential target region detection network, which leverages supervised training along with feature alignment to transfer knowledge from the multi-modal large model to enhance performance. Then, region localization is refined by smoothing the confidence map. Finally, a lightweight model is applied to high-resolution cropped images for precise defect, avoiding issues caused by excessive scaling of objects. To verify the effectiveness of this method, a small fitting defect detection dataset and conducted experiments are constructed. Results show that the proposed algorithm improves precision and recall values by an average of about 10%, significantly outperforming methods that directly apply detection models on full images. Experimental results demonstrate that this method has good comprehensive performance and can effectively handle the challenge of defect detection in tiny components within high-resolution images.