AdaMorphNet: Morphology-Aware Weakly-Supervised Defect Localization for Nondestructive Radiographic Weld Inspection

Abstract

High annotation costs and the difficulty of localizing fine-grained linear defects limit the deployment of deep learning in automated nondestructive evaluation (NDE) of welds. Weakly-Supervised Object Localization (WSOL) reduces labeling effort, but standard Class Activation Map (CAM) approaches produce diffuse heatmaps that do not resolve the morphology of thin welding discontinuities. We introduce AdaMorphNet, a framework pairing multi-scale CAM fusion with defect-specific geometric constraints drawn from radiographic inspection standards. In place of uniform thresholding, the model refines coarse activations with one constraint per defect class: topology-preserving thresholding for cracks, an anisotropic height limit for lack of penetration, and distribution-adaptive clustering for porosity. On the RIAWELC benchmark AdaMorphNet reaches 99.96% classification accuracy. Against 1,840 manually verified ground-truth boxes it attains 56.81% mean IoU and 97.88% Pointing Accuracy, exceeding a standard Grad-CAM baseline by 20.81 percentage points; an architectural ablation attributes 8.07 points of mean IoU to multi-scale fusion over single-scale CAM under identical post-processing. The pipeline localizes 37.1 patches of \({224 \times 224}\) pixels per second on consumer hardware, which supports near-real-time pre-screening at an estimated 1.8 radiographs per second under a two-stage triage schedule, without dense pixel-level supervision.