Abstract: To address challenges in laser soft soldering joint defect detection, such as limited sample availability, high sensitivity to overlapping bounding boxes, small target size, and complex fine-grained features, a novel small-sample detection approach is proposed. This method integrates Generative Adversarial Network (GAN)-based data augmentation with deep transfer learning. First, a PCB solder joint defect dataset is constructed using an integrated visual acquisition system. Then, an Auxiliary Classifier GAN (ACGAN) enhanced with hybrid attention mechanisms and spectral normalization is developed for image augmentation. A YOLOv5-TL transfer learning detection model is built, and the transfer strategy is optimized accordingly. Experimental results demonstrate that the proposed data augmentation improves the detection accuracy of the non-transfer model by approximately 2.9%, while the transfer learning strategy provides an additional gain of about 5.7%. When combined, an overall improvement of around 11.1% is achieved. Although data augmentation slightly slows convergence, transfer learning significantly accelerates it, and together they improve convergence speed by about 12.5%. The proposed method has been successfully deployed in a laser soft soldering system, enabling real-time online defect detection with high accuracy and system stability, thus meeting practical application requirements.