A novel method for three-dimensional liquid surface topography measurement: Combining binocular vision with deep learning-based digital image correlation

Abstract

Liquid surface topography measurement is an urgent need for research on the nonlinear sloshing dynamics of liquids. However, it poses challenges due to liquids' inherent properties like high fluidity, transparency, and specular reflection. The paper proposes a novel method for three-dimensional liquid surface topography measurement combining binocular vision and deep learning-based digital image correlation (Deep-DIC). The method employs a binocular vision system for the acquisition of speckle projection images of the liquid surface. A UNet-based Deep-DIC model is constructed and datasets containing low-contrast, out-of-focus samples are generated for model training in a targeted manner, considering the limitations of liquid measurements. The Deep-DIC-driven spatially dense stereo matching is then conducted and combined with the binocular stereo vision imaging model to achieve three-dimensional liquid surface topography measurements. The method's effectiveness is validated through experimental investigations under static and harmonic excitation conditions. The results of static experiments showcase the ability to reconstruct liquid surfaces with varying heights. The proposed method achieves a maximum mean off-plane error of 1.226 mm with high-quality speckle projection and only 2.342 mm even with low-quality speckle projection, demonstrating clear superiority over traditional DIC. Harmonic excitation experiments further validate the method's capability in capturing dynamic liquid surface vibrations, with results closely aligning with those of finite element method. The relative error between the measured main frequency and the true excitation frequency is only 2.75 %. This study provides a new perspective and has the potential to inform research on liquid sloshing in industrial liquid storage systems.