|
|
 |
| |
|
|
|
Advanced Sustainable Engineering Innovation of Hybrid Learning for Single-Image Depth Estimation |
|
|
|
PP: 175-182 |
|
|
doi:10.18576/ijtfst/150204
|
|
|
|
Author(s) |
|
|
|
Divya Kiran,
Ugra Mohan Roy,
|
|
|
|
Abstract |
|
|
| Advanced thin-film engineering is important in modern electronics, energy systems, optical devices, and smart manufacturing. Accurate surface characterization and defect analysis are key to making sure these technologies work well and remain reliable. This study presents a Hybrid Learning Framework for Single-Image Depth Estimation (SIDE) to support non- contact inspection and three-dimensional surface reconstruction of thin-film structures. The proposed framework addresses the ill-posed nature of monocular depth estimation by integrating probabilistic uncertainty modeling with deep learning-based metric refinement. Specifically, a Bayesian Network (BN) is employed to capture geometric priors and quantify heteroscedastic uncertainty, while a Supervised Convolutional Neural Network (CNN) performs depth refinement using a multi-objective loss function. Experimental evaluation on the NYU Depth V2 benchmark achieved an Absolute Relative Error of 0.068, demonstrating superior performance compared with conventional deterministic approaches. The integration of BN and CNN models enables the generation of confidence- aware depth maps that improve the reliability of surface characterization and structural assessment in thin-film engineering applications. The proposed hybrid framework offers a cost-effective and scalable solution for intelligent inspection, defect detection, and digital manufacturing environments, contributing to the advancement of next-generation thin-film engineering innovations. |
|
|
|
|
|
|
|
