Dynamic zoom-in detection of exfoliated two-dimensional crystals using deep reinforcement learning

TL;DR

This paper introduces a deep reinforcement learning-based method for efficiently detecting exfoliated two-dimensional crystals in high-resolution images by dynamically zooming into regions of interest, outperforming baseline detection methods.

Contribution

It presents a novel combination of deep reinforcement learning and object detection for targeted crystal search, improving efficiency and accuracy in identifying microflakes.

Findings

Outperformed baseline detection methods

Efficiently searches for various crystal types

Analyzed the search efficiency of the RL agent

Abstract

Owing to their tunability and versatility, the two-dimensional materials are an excellent platform to conduct a variety of experiments. However, laborious device fabrication procedures remain as a major experimental challenge. One bottleneck is searching small target crystals from a large number of exfoliated crystals that greatly vary in shapes and sizes. We present a method, based on a combination of deep reinforcement learning and object detection, to accurately and efficiently discover target crystals from a high resolution image containing many microflakes. The proposed method dynamically zooms in to the region of interest and inspects it with a fine detector. Our method can be customized for searching various types of crystals with a modest computation power. We show that our method outperformed a simple baseline in detection tasks. Finally, we analyze the efficiency of the deep reinforcement learning agent in searching crystals. Codes are available at \url{https://github.com/stephandkim/detect_crystals}.