Anomaly detection is a critical problem in the manufacturing industry. In many applications, images of objects to be analyzed are captured from multiple perspectives which can be exploited to improve the robustness of anomaly detection. In this work, we build upon the deep support vector data description algorithm and address multi-perspective anomaly detection using three different fusion techniques, i.e., early fusion, late fusion, and late fusion with multiple decoders. We employ different augmentation techniques with a denoising process to deal with scarce one-class data, which further improves the performance (ROC AUC ). Furthermore, we introduce the dices dataset, which consists of over 2000 grayscale images of falling dices from multiple perspectives, with 5% of the images containing rare anomalies (e.g., drill holes, sawing, or scratches). We evaluate our approach on the new dices dataset using images from two different perspectives and also benchmark on the standard MNIST dataset. Extensive experiments demonstrate that our proposed multi-perspective approach exceeds the state-of-the-art single-perspective anomaly detection on both the MNIST and dices datasets. To the best of our knowledge, this is the first work that focuses on addressing multi-perspective anomaly detection in images by jointly using different perspectives together with one single objective function for anomaly detection.
9 July 2021 in The International Journal of Advanced Manufacturing Technology
In industrial manufacturing of bright steel rods, one important quality factor is the straightness or straightness deviation. Depending on the application, deviations of less than 0.1 mm per meter rod length are desired and can be reached with state-of-the-art manufacturing equipment. Such high-quality requirements can only be guaranteed with continuous quality control. Manual straightness measurements conducted offline using a dial gauge provide accurate results on single positions of the rod. We propose a contactless, optical measurement technique based on laser scanning which has the potential to be used inline during production to inspect all rods over the entire length. Only for calibration of the system the rod needs to be turned around its axis. For the measurement of straightness deviation, it is not required to turn the rod. The method is based on evaluating the intensity signal of the reflected laser radiation against the scan angle. It is shown that in combination with an accurate calibration, this signal can be used to determine the rod’s deviation from a straight rod. We explain the measurement and calibration principle as well as data evaluation. We present the experimental setup and first measurement results on a single position on several samples. For a homogeneous sample surface and neglecting laser drift, accuracy and precision were determined to be in the range of 10–20 μm. We discuss the working principle of a potential inline system.
Nodal Aberration Theory (NAT) was discovered in 1977 and was last presented at this conference 25 years ago. This paper presents recent discoveries and why this theory is gaining in relevance for the future.
Am Institut für Umformtechnik der Universität Stuttgart wurde eine automatisierte Messanlage in eine Kaltmassivumformungs-Produktionslinie integriert. Das Projekt zielt darauf ab, Bauteile durch einen digitalen „Fingerabdruck“ zu identifizieren, Prozess- und Bauteildaten zu erfassen und sie mit individuellen Pressteilen zu verknüpfen. Die Schlüsselkomponente ist eine Prüfkugel des Fraunhofer IPM, die während des freien Falls des Pressteils Fotos aufnimmt und sie für Kontrolle, Identifikation und Fehlererkennung nutzt. The Institute for Metal Forming Technology at the University of Stuttgart collaborated with partners to automate measurement in a cold forging production line. The project aims for marker-free component identification via „Fingerprint Track & Trace“, process and component data measurement, and data association with individual components. A key component is Fraunhofer IPM‘s inspection sphere, capturing photos of falling pressed parts for dimensional accuracy checks, error detection, and fingerprint determination.
Anomaly detection is a critical problem in the manufacturing industry. In many applications, images of objects to be analyzed are captured from multiple perspectives which can be exploited to improve the robustness of anomaly detection. In this work, we build upon the deep support vector data description algorithm and address multi-perspective anomaly detection using three different fusion techniques, i.e., early fusion, late fusion, and late fusion with multiple decoders. We employ different augmentation techniques with a denoising process to deal with scarce one-class data, which further improves the performance (ROC AUC $= 80\%$). Furthermore, we introduce the dices dataset, which consists of over 2000 grayscale images of falling dices from multiple perspectives, with 5\% of the images containing rare anomalies (e.g., drill holes, sawing, or scratches). We evaluate our approach on the new dices dataset using images from two different perspectives and also benchmark on the standard MNIST dataset. Extensive experiments demonstrate that our proposed {multi-perspective} approach exceeds the state-of-the-art {single-perspective anomaly detection on both the MNIST and dices datasets}. To the best of our knowledge, this is the first work that focuses on addressing multi-perspective anomaly detection in images by jointly using different perspectives together with one single objective function for anomaly detection.