Efficient automated high dynamic range 3D measurement via deep reinforcement learning

Abstract

High dynamic range 3D measurement technology, utilizing multiple exposures, is pivotal in industrial metrology. However, selecting the optimal exposure sequence to balance measurement efficiency and quality remains challenging. This study reinterprets this challenge as a Markov decision problem and presents an innovative exposure selection method rooted in deep reinforcement learning. Our approach’s foundation is the exposure image prediction network (EIPN), designed to predict images under specific exposures, thereby simulating a virtual environment. Concurrently, we establish a reward function that amalgamates considerations of exposure number, exposure time, coverage, and accuracy, providing a comprehensive task definition and precise feedback. Building upon these foundational elements, the exposure selection network (ESN) emerges as the centerpiece of our strategy, acting decisively as an agent to derive the optimal exposure sequence selection. Experiments prove that the proposed method can obtain similar coverage (0.997 vs. 1) and precision (0.0263 mm vs. 0.0230 mm) with fewer exposures (generally 4) compared to the results of 20 exposures.