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

Area of Science:

• Metrology
• Computer Vision
• Machine Learning

Background:

• High dynamic range 3D measurement is crucial for industrial applications.
• Optimizing exposure sequences for efficiency and quality in 3D metrology is a significant challenge.

Purpose of the Study:

• To develop an innovative exposure selection method for high dynamic range 3D measurement using deep reinforcement learning.
• To address the challenge of balancing measurement efficiency and quality in 3D metrology.

Main Methods:

• Reinterpreting exposure selection as a Markov decision problem.
• Developing an exposure image prediction network (EIPN) to simulate a virtual environment.
• Establishing a reward function incorporating exposure number, time, coverage, and accuracy.
• Utilizing an exposure selection network (ESN) as an agent to derive optimal exposure sequences.

Main Results:

• The proposed deep reinforcement learning method achieves comparable coverage (0.997) and precision (0.0263 mm) to traditional methods.
• The method significantly reduces the number of required exposures, typically using 4 exposures compared to 20.
• Experimental results demonstrate the effectiveness of the ESN in selecting optimal exposure sequences.

Conclusions:

• Deep reinforcement learning offers an effective solution for optimizing exposure sequences in high dynamic range 3D measurement.
• The developed method enhances measurement efficiency without compromising quality, offering practical benefits for industrial metrology.