In the past few years, object detection technology has developed rapidly and been widely used in life. However, these techniques as closed-set detectors can only detect the categories existing in the training data, and there is a problem of poor transfer effect in small-scale datasets. Therefore, Grounding DINO uses large language models’ (LLMs) powerful text semantic encoding to extend closed-set detectors to open-set, achieving better results on small-scale datasets. We improved it with novel Text-guided Deformable Attention mechanism for large scale variation in transmission inspection specially. As a result, we achieved the optimal transfer results of 7.3% mAP0.5 higher than other models on average in transmission inspection.