A novel surface denoising approach based on deep learning for freeform structured surface in metrology

Noise produced in manufacturing and measuring process are often high-frequency components and outliers, these will cause the inaccuracy of feature determination. Surface denoising is a crucial metrological operation in surface characterisation to suppress noise from areal surface. It usually called filtration for processing profile data. ISO standards filters such as Gaussian filters and wavelets filters have been proved successful on stochastic surfaces. PDE diffusion filter has been subsequently raised to address edge distortion in structured surface denoising. The main issue is that they all require manual parameter tuning for individual surface. This process costs time and relies on expert-knowledge of users. To date, the deep learning techniques are becoming dominated in image processing tasks including denoising, object detection and classification, which would also have great potential to benefit surface metrology. This paper proposed a deep learning-based method applying a classic deep Convolutional Neural Network to perform surface denoising task. The model is trained on a small training dataset of freeform structured surface measurements. The experimental results show that retrained neural network can automatically suppress unknown noises and outliers of the surfaces, meanwhile well retain the geometry of structures on the surface. The major contribution is that we newly apply a deep convolutional neural network to replace traditional filters and achieve automatic surface denoising. It can output denoising results within average one second, which shows a high application value for constructing future smart metrology system. The training process can be efficiently implemented on GPU at a low cost.