Deep Learning-Based Defect Segmentation in PAUT B-Scan Images for Nondestructive Evaluation of Metallic Blocks
Le Khuong Phan, Dinh Tuan Nguyen, Thi Thu Ha Vu, Tan Hung Vo, Anh Kiet Nguyen, Jaeyeop Choi, Jae Sung Ahn, Sudip Mondal, Junghwan Oh
Metallic blocks and components are indispensable across the aerospace, energy, and heavy-engineering industries, where undetected internal flaws such as cracks, voids, and inclusions may precipitate catastrophic structural failure. Reliable nondestructive evaluation (NDE) is essential to ensure their integrity and operational safety. Among the available NDE techniques, phased array ultrasonic testing (PAUT) has emerged as one of the most accessible and widely adopted, by virtue of its rapid scanning, electronic beam steering, and capacity to image subsurface defects without disassembly. However, the interpretation of PAUT B-scan images remains hindered by background reflections, material-dependent echo characteristics, and substantial variability in defect size. In this work, a fine-tuned encoder–decoder deep learning network is proposed for the automated segmentation of internal defects in PAUT B-scan images of metallic block specimens. The network couples a ResNet50 encoder with a shallow detail stem, a multi-scale feature fusion module, and a detail refinement block, designed to preserve small defect echoes and sharpen weak defect boundaries characteristic of internal flaws. The proposed approach was compared with five state-of-the-art segmentation architectures, namely FCN, PSPNet, DeepLabv3+, HRNet-OCR, and SegFormer, as well as a conventional Otsu-thresholding baseline representing standard PAUT screening practice. Experimental results demonstrate that the proposed network attained the highest Dice score of 0.7964, defect intersection-over-union of 0.6617, and precision of 0.7094 among all evaluated models, while the Otsu baseline yielded the lowest scores, confirming the benefit of learned segmentation over fixed amplitude thresholding. These findings indicate that the proposed network achieves a favorable trade-off between defect localization accuracy and false-positive suppression, underscoring its potential for reliably segmenting internal defects in PAUT B-scan imaging.