Content-based image retrieval for industrial material images with deep learning and encoded physical properties

Abstract Industrial materials images are an important application domain for content-based image retrieval. Users need to quickly search databases that exhibit similar appearance, properties, and/or features reduce analysis turnaround time and cost. The in this study 2D of millimeter-scale rock samples acquired at micrometer resolution with light microscopy or extracted from 3D micro-CT scans. Labeled expensive time-consuming acquire thus typically only available the tens thousands. Training a high-capacity deep learning (DL) model scratch is therefore not practicable due data paucity. To overcome “few-shot learning” challenge, we propose leveraging pretrained common DL models conjunction transfer learning. “similarity” industrial subjective assessed by human experts based on both visual appearance physical qualities. We have emulated human-driven assessment process via physics-informed neural network including metadata measurements loss function. present novel architecture combines Siamese networks function integrates classification regression terms. trained similarity (classification), prediction (regression). For efficient inference, use highly compressed feature representation, computed offline once, database query image. Numerical experiments demonstrate superior retrieval performance our new compared other custom-feature-based approaches.