VLM-TD: A Visual Language Model for Transmission Defects with Integrated Link Attention

Traditional transmission line inspection, which relies on manual recording of fault information, is prone to ambiguity. The semantics generated by general image description models suffer from poor professional expression and low accuracy, failing to adequately meet the needs of transmission line defect detection. This paper proposes a Link Attention (LA) mechanism, which improves the visual-language model through dual-axis extraction. Integrated into the Vision Transformer (ViT) architecture as a visual encoder, it enhances the model's deep understanding of power features and its ability to represent the relationship between defects and power equipment. To strengthen the mining capability of visual features and the collaborative representation of multimodal features, and to mitigate the insufficient coverage of small sample distributions, the crossattention in the Bert architecture is improved to Dual Cross Attention (DCA). Further training and optimization through the introduction of a transfer learning strategy yield the Visual Language Model for Transmission Defects (VLM-TD). Experimental results show that VLM-TD, compared to the baseline model, achieves improvements of $8.4 \%, 7.3 \%$, and 3.2 % in BLEU-1, BLEU-4, and SPICE metrics, respectively, in the constructed transmission defect image-text pair dataset, and an average improvement of 4 % in COCO and Flickr datasets. The accuracy of generated descriptions reaches 84.9 %, providing a new solution for transmission line defect detection.