Fraunhofer-Institut für Produktionsanlagen und Konstruktionstechnik IPK

Filters

14
14
14
Reset filters

Settings
Now showing 1 - 10 of 14
  • Publication
    Towards anomaly detection: A feasibilty study for quality control in productions of nonwovens
    ( 2024)
    Lehr, Jan
    ;
    Pape, Martin
    ;
    Nguyen Hoang, Viet
    ;
    Philipps, Jan
    ;
    Krüger, Jörg
    In the face of the upcoming Corona pandemic in 2020, which continues to this day, adequate protection for all citizens has been very important and continues to be essential. An easy way to minimize the risk of infection in public, but poorly ventilated places, are FFP masks. The basic material from which FFP masks are made is nonwoven. The quality of the produced nonwoven is the basis for the correct function and thus the protection against infection. The present research work investigates methods of computer vision to detect contaminations and damages on the nonwoven. Both supervised and semi supervised methods are evaluated. For this work, an inspection system consisting of two separate acquisition systems was developed for image data acquisition. The first system is suitable for the visible wavelength and has a theoretical resolution of 2.4 μm per pixel. The second camera system is designed for the near infrared range and has a theoretical resolution of 5.5 μm per pixel. The acquisition system collects an image data set comprising 1,760 images with 920 images of defect-free nonwoven samples and 840 images of defective nonwoven samples. First, the wavelength range suitable for optical inspection of uncoated nonwoven samples is investigated. A further investigation dealt with the question of whether the reflected light or transmitted light method is more suitable for optical inspection. Finally, coated nonwoven is also inspected using the reflected light method. An investigation using the transmitted light method is not possible, as the material is almost non-transparent. Despite the small amount of data, very good results were achieved. Machine learning methods from the field of image processing are usually classified as deep learning. This means that the large network architectures require very large amounts of data in order to learn complex patterns. Publicly available datasets for method evaluation typically consist of over 1,000 images per class or defect. The nonwoven samples provided in this work and the resulting image database is about a factor of six smaller than is actually intended for the methods used. In this work it has been possible to achieve inspection accuracies of 97.5 %.
  • Publication
    Insights of Anomaly Detection: How does Polluted Training Data Influence Performance?
    ( 2024)
    Lehr, Jan
    ;
    Pape, Martin
    ;
    Günther, Samuel
    ;
    Krüger, Jörg
    Anomaly detection is one of the most popular fields for computer vision in industrial applications. The idea of training machine learning only on defect-free objects saves enormous amounts of integration effort. The state of the art shows that current methods on public data sets (e.g. MVTec AD data set [1]) have already solved the problem with AUROC segmentation scores of more than 99 %. In real-world applications training data is not as "clean" as in public data sets. This work investigates the changes in detection performance when outliers end up in the training data. For this purpose, the training data is enriched step by step with images of defective objects. The AUROC score and the anomaly score is used as a quality criterion for performance measurement. We show that state of the art methods can be very robust, but that in some scenarios a draw down of 15 percentage points is possible.
  • Publication
    Image-Based Incremental Learning for Part Recognition of Used Automotive Cores in Reverse Logistics
    ( 2024)
    Briese, Clemens
    ;
    Chavan, Vivek Prabhakar
    ;
    Schlüter, Marian
    ;
    Lehr, Jan
    ;
    Kröger, Ole
    This paper presents a study on image-based incremental learning for part recognition of used automotive parts, also known as cores. The use of Machine Learning (ML) in the recognition of used parts has proven to be effective in suggesting Original Equipment Number (OEN) based on images and logistics data of a core. This leads to a four-eye process where the worker and ML interact through an assistance system. In reverse logistics, the spectrum of parts handled is constantly changing, making it difficult to have a "complete" image or sensor-based data set. The study focuses on the ramp-up phase of an ML implementation project in a real-world automotive core sorting station. There are two stations equipped with sensors such as RGB cameras. The sorted parts were acquired over a period of one year. Incremental learning was employed to cope with the growing dataset and the growing number of classes to be identified without retraining a model from scratch. Open source and state-of-the-art incremental ML learning methods such as POD-Net and Foster were tested against the common joint training approach used for most benchmarks in computer vision. The best-fitting open-source method for this problem was identified as POD-Net used with a self-supervised pretrained ResNet50. For the ramp-up of an ML-based core recognition a combination of incremental learning and joint training was found to be useful. It starts learning from a small number of digitized parts (14 classes), while maintaining a high recognition accuracy rate throughout the year, with a final class count of 100 (an increase of approx. 600%), which is a subset of a real application problem. The results of this study show that the proposed method is efficient, plastic, and energy-saving. Thus, it is a promising approach for the recognition of used automotive parts.
  • Publication
    Limitations of Anomaly Detection: Beyond which Size Defects can be Reliably Recognized
    ( 2024)
    Lehr, Jan
    ;
    Pape, Martin
    ;
    Philipps, Jan
    ;
    Scholler, Felix
    ;
    Krüger, Jörg
    Anomaly detection is one of the most popular fields for computer vision in industrial applications. The idea of training machine learning only on defect-free objects saves enormous amounts of integration effort. The state of the art shows that current methods on public data sets (e.g. MVTec AD data set) have already solved the problem with AUROC segementations scores of more than 99 %. But how accurate are these methods really? In this paper, one current method from the field of supervised learning and anomaly detection is evaluated on two problems. Each problem contains a defect pattern that grows in 11 steps. This work shows that the defect is already reliably detected from a relative size of 0.03 % of the pixels in the image.
  • Publication
    Image-based recognition of surgical instruments by means of convolutional neural networks
    ( 2023)
    Lehr, Jan
    ;
    Kelterborn, Kathrin
    ;
    Briese, Clemens
    ;
    Schlueter, Marian
    ;
    Kroeger, Ole
    ;
    Krüger, Jörg
    Purpose: This work presents a novel camera-based approach for the visual recognition of surgical instruments. In contrast to the state of the art, the presented approach works without any additional markers. The recognition is the first step for the implementation of tracking and tracing of instruments wherever they are visible and could be seen by camera systems. Recognition takes place at item number level. Surgical instruments that share the same article number also share the same functions. A distinction at this level of detail is sufficient for most clinical applications. Methods: In this work, an image-based data set with over 6500 images is generated from 156 different surgical instruments. Forty-two images were acquired from each surgical instrument. The largest part is used to train convolutional neural networks (CNNs). The CNN is used as a classifier, where each class corresponds to an article number of the surgical instruments used. Only one surgical instrument exists per article number in the data set. Results: With a suitable amount of validation and test data, different CNN approaches are evaluated. The results show a recognition accuracy of up to 99.9% for the test data. To achieve these accuracies, an EfficientNet-B7 was used. It was also pre-trained on the ImageNet data set and then fine-tuned on the given data. This means that no weights were frozen during the training, but all layers were trained. Conclusion: With recognition accuracies of up to 99.9% on a highly meaningful test data set, recognition of surgical instruments is suitable for many track and trace applications in the hospital. But the system has limitations: A homogeneous background and controlled lighting conditions are required. The detection of multiple instruments in one image in front of various backgrounds is part of future work.
  • Publication
    Automated Optical Inspection Using Anomaly Detection and Unsupervised Defect Clustering
    ( 2020)
    Lehr, Jan
    ;
    Sargsyan, Arlik
    ;
    Pape, Martin
    ;
    Philipps, Jan
    ;
    Krüger, Jörg
    Neural networks have proven to be extraordinarily successful in many computer vision applications. But the approaches used to train neural networks require large datasets of annotated images, which requires a solid amount of human time to prepare those datasets. To facilitate the adoption of machine learning based technologies in industrial computer vision applications, this paper presents a two-step unsupervised learning approach for anomaly detection with further defect clusterization. In the first stage, the defects are not explicitly learned, but are interpreted as an anomaly or novelty based on the dataset of defect-free samples. In a second stage, the anomalies detected in the first stage are clustered in unsupervised manner and classified into meaningful categories by experts with process knowledge (e.g. critical or non-critical defect). This paper presents a first small dataset containing one industrial object with a complex shape. The object is made of aluminiu m and is shown both free of defects and defective. Based on this, recommendations are given for an acquisition setup for a large, extensive dataset. Most of the existing papers are studying the approaches for uniform surface (texture) inspection. The specifics of this research is to identify defects on rigid bodies, which exhibit highly non uniform texture in the image. State of the art methods were evaluated and improved to increase the classification accuracy. With a fine-tuned ResNet-18 it was possible to achieve 100% accuracy for defective and defect-free images.
  • Publication
    Decentralised identification of used exchange parts with a mobile application
    ( 2020)
    Lehr, Jan
    ;
    Schlüter, Marian
    ;
    Krüger, Jörg
    Sustainable product development and use requires an extended life cycle of used and defective mechanical parts. Remanufacturing saves resources and helps the industry to utilise the product more efficiently. Reverse logistics is one of the most important challenges towards efficient remanufacturing. To improve this process, we propose an on-site part identification at the workshops. A fast on-site identification is essential for assisting repair shop personnel and saving time on searching for the right spare parts. Based on images taken by a mobile device our application provides various machine vision services, e.g., visual identification of used parts, already successfully tested in a sorting facility for remanufacturing parts. The mobile application provides a robust visual identification for different environments. We show that enhancing data for machine vision approaches with images from decentral sensors, i.e., mobile devices, leads to an improved identification accuracy.