Weld seam type confirmation is a key part of intelligent integrated welding to cope with adjustable schemes varying with the seam morphology. However, existing works are mainly based on qualitative joint description (QJD) and machine learning classification (MLC), which entail high costs in balancing multiple given parameters, acquiring sufficient data and generalizing for new seam types. In this article, taking the advantages of few-shot learning, we propose a full-cycle data purification strategy (DPS) to identify seam types. First, five typical weld seams are mapped under three laser stripe patterns to acquire the raw samples. Then, a series of compliance processing algorithms are performed with reference to the target dataset to build a scenario dataset: purified weld seam stripe mapping (P-WSSM). With a few-shot learning pre-model, a positive weld seam classification result is obtained based on a small-volume dataset and analyze several critical factors. In addition, to compensate the application drawbacks caused by the wide randomness of low-volume datasets, compliance distance based on integrated grayscale co-occurrence matrix (I-GLCM) is introduced to quantitatively measure P-WSSM and the target dataset. Ultimately, experiments show that the dual-line stripe mapping achieves 85.4% recognition accuracy for 5-way 5-shot, which is a competitive multi-type weld seam classification via few-shot learning.