Toward Automated Penetrant Testing: Monocular Vision Target Localization With Optimizable Weighted Latent Representation Learning

Penetrant testing (PT) is a critical nondestructive inspection method in industry. While robotic automation improves PT repeatability and safety, it demands accurate and scalable spatial measurement. Traditional high-precision instrumentation lacks scalability and vision-based methods struggle to balance computational efficiency and the preservation of spatially critical information. This article develops a monocular vision target localization method for automated PT systems, focusing on deterministic spatial-cue preservation during feature compression and motion-consistent depth refinement. The proposed approach integrates a feature-based weighting module into a depth estimation network, selectively preserving essential spatial cues while suppressing irrelevant details to enhance robustness under constrained representation lengths. A motion-aware depth optimization step then enforces interframe consistency by leveraging known manipulator motions as a strong geometric prior. Experiments on standard indoor benchmarks show consistent accuracy gains. More importantly, evaluations on a novel simulated PT dataset show a significant performance improvement (∼27.2% root-mean-squared error (RMSE) reduction) over representative baselines. A full workflow demonstration further confirms its operational efficacy and measurement reliability in real-world PT scenarios.