3D Measurement Using a Single Image for Smart Manufacturing of Microscopic Products in a Ceramic Powder Pressing Process

Abstract

Owing to the development of artificial intelligence, smart manufacturing has extensively been investigated, which has rapidly increased manufacturing productivity. The most prevalent smart manufacturing application using machine learning is product quality enhancement through surface and dimensional defect detection. Various sensors, such as stereoscopic, 2D, or 3D lasers as well as structured light sensors, have been used to perform precise thickness measurements. These sensors are expensive and they require a large space for installation in a factory, which consumes time and energy. Moreover, micro-scale dimensional measurement equipment is much more expensive and easily damaged by environmental changes such as temperature and humidity. In the beginning, we introduce a single image-based 3D measurement system using depth estimation that makes dimensional anomaly detection easy and fast. This is important for green technology because the proposed method can reduce condemned or inferior goods by detecting anomalies early and save raw materials and energy consumption. The existing method is deep learning based depth estimation. However, it is difficult for conventional depth estimation methods to predict 3D measurement directly because of the difference between depth estimation and 3D measurement. That is because inputs are discontinuous dimensional values with homogeneous textures. Conventional depth estimation is a regression task that assumes inputs are continuous. To circumvent this problem, we propose a magnifier loss. In addition, to overcome the object's homogeneity and meet the micro-scale dimensional precision requirement, this paper proposes a novel magnifier transformation function that magnifies the homogeneous textures and micro-scale dimensional value changes. This method achieves better quantitative performance than a conventional computer vision method, stereo-matching approaches, and even a structured light sensor, which is known to be one of most accurate and expensive sensors for a 3D measurement system.