To address the missed detection of small targets, spatial sensitivity of key weld regions, and blurred defect boundaries in T-type weld defect detection, an improved YOLO11s-based method is proposed. A Coordinate Attention (CA) module is introduced to enhance the spatial feature representation of key weld regions. The SIoU loss is employed to optimize bounding-box regression, and a P2 detection layer is added to the detection head to supplement shallow high-resolution details. Experiments were conducted using a public T-type weld defect dataset and an expanded training set. The proposed model achieved a precision of 87.8%, a recall of 78.0%, and an mAP@0.5 of 86.4%, outperforming YOLO11s by 5.6, 8.2, and 3.8 percentage points, respectively, while increasing the number of parameters only from 9.43 M to 9.60 M. Under low-contrast, high-noise, blurred, and mixed-degradation conditions, the proposed model improved mAP@0.5 by 8.53–18.95 percentage points compared with YOLO11s. With FP16 precision and an input size of 960×960, the model achieved an end-to-end inference speed of 66.20 FPS on an RTX 4060 Ti. The proposed method improves detection accuracy and adaptability to complex imaging conditions while maintaining favorable real-time performance, providing a reference for intelligent weld quality inspection.