To address the issues in ultrasonic testing of pipeline welds, including reliance on manual expertise, insufficient generalization capability of deep learning models, and scarcity of annotated samples, this paper proposes an intelligent identification method based on Gramian Angular Field (GAF) and an improved Transformer-Faster R-CNN. By converting one-dimensional ultrasonic time-series signals into two-dimensional feature images via GAF encoding, the method preserves the temporal dependencies and spatial patterns of defects. A hybrid backbone network integrating ResNet-101 and a Transformer encoder is introduced to combine local feature extraction with global contextual awareness. Additionally, a spatial attention module is embedded in the Region Proposal Network (RPN) to enhance the mod-el's focus on discriminative defect features. Experimental results demonstrate that, compared to models such as ResNet, VGG, Vision Transformer (Vit), and the original Faster R-CNN, the proposed method achieves significant improvements in accuracy, precision, recall, and F1-score, reaching up to 96.94%, 97.21%, 96.96%, and 96.90%, respectively. It exhibits superior identification accuracy and robust-ness, particularly for small-sized and complex-shaped defects. This approach provides an efficient solution for the automated and intelli-gent identification of pipeline weld defects.