Transfer-learning-based cross-scale pearlite grain segmentation for low-carbon pipeline steels

Accurate segmentation of pearlite grains across multiple length scales is essential for evaluating the mechanical properties of low-carbon pipeline steels. Traditional manual methods are labor-intensive and suffer from poor repeatability, whereas existing deep-learning models exhibit significant accuracy drops when confronted with variable scales, etching contrasts, or imaging conditions. To address these challenges, this paper introduces a transfer-learning framework based on a Residual U-Net architecture, originally pre-trained on thousands of ferrite–pearlite micrographs via the SPFCA semi-automatic labeling algorithm. Two complementary transfer strategies – Frozen-encoder and Fine-tuning-all – are systematically investigated. A carefully curated dataset, derived from welded L245 N pipeline steel and partitioned into base-metal (coarse grains) and heat-affected-zone (fine grains) subsets, is employed for validation. Quantitative evaluation using precision, recall, and F1-score demonstrates that both transfer strategies substantially outperform the baseline model trained from scratch. Specifically, the Frozen strategy excels when target scales resemble the pre-training domain (F1 = 0.911 compared to 0.878), whereas Fine-tuning proves superior for smaller-scale targets (F1 = 0.887 compared to 0.791), confirming the adaptability of transfer learning to cross-scale segmentation tasks. The proposed framework significantly reduces annotation costs and provides a robust, scalable solution for automated metallographic quality control.