Depth Data Denoising in Optical Laser Based Sensors for Metal Sheet Flatness Measurement: A Deep Learning Approach

Area of Science:

• Materials Science
• Computer Vision
• Metrology

Background:

• Surface flatness is crucial for metal sheet quality control.
• Laser-based optical sensors offer accurate surface reconstruction but are susceptible to specific noise.
• Effective image denoising is essential for reliable flatness measurements.

Purpose of the Study:

• To develop a deep learning architecture for denoising range images from laser-based sensors in metal sheet production.
• To address specific noise patterns caused by mechanical vibrations during manufacturing.
• To enhance the accuracy of surface flatness assessment systems.

Main Methods:

• A novel convolutional blind residual denoising network (CBRDNet) was proposed.
• CBRDNet integrates noise estimation and removal modules using semantic convolutional neural networks.
• The network was validated on synthetic and real-world noisy range image data.

Main Results:

• CBRDNet demonstrated superior performance over traditional 1D/2D filtering methods.
• It outperformed state-of-the-art CNN-based denoising techniques.
• Error reduction was up to 15% compared to traditional methods and 3-10% compared to other deep learning methods.

Conclusions:

• The proposed CBRDNet effectively removes critical noise from range images in metal sheet production.
• This deep learning approach significantly improves the accuracy of flatness measurements.
• CBRDNet offers a robust solution for enhancing quality control in the metal industry.