Eignung von Convolutional-Auto-Encodern zur automatisierten Fehlererkennung bei Serienbauteilen am Beispiel von Airbag-Generatoren
Rostock, Univ., Master Thesis, 2019
Underwater Color Restoration Using U-Net Denoising Autoencoder
Proceedings of the 11th International Symposium Image and Signal Processing and Analysis
International Symposium on Image and Signal Processing and Analysis (ISPA) <11, 2019, Dubrovnik, Croatia>
Visual inspection of underwater structures by vehicles, e.g. remotely operated vehicles (ROVs), plays an important role in scientific, military, and commercial sectors. However, the automatic extraction of information using software tools is hindered by the characteristics of water which degrade the quality of captured videos. As a contribution for restoring the color of underwater images, Underwater Denoising Autoencoder (UDAE) model is developed using a denoising autoencoder with U-Net architecture. The proposed network takes into consideration the accuracy and the computation cost to enable realtime implementation on underwater visual tasks using end-toend autoencoder network. Underwater vehicles perception is improved by reconstructing captured frames; hence obtaining better performance in underwater tasks. Related learning methods use generative adversarial networks (GANs) to generate color corrected underwater images, and to our knowledge this paper is the first to deal with a single autoencoder capable of producing same or better results. Moreover, image pairs are constructed for training the proposed network, where it is hard to obtain such dataset from underwater scenery. At the end, the proposed model is compared to a state-of-the-art method.
Underwater Image Enhancement Using Autoencoders
Rostock, Univ., Master Thesis, 2019
Visual inspection of underwater structures by vehicles, e.g. remotely operated vehicles (ROVs), plays an important role in scientific, military, and commercial aspects. However, the automatic extraction of information using software tools is hindered by the optical characteristics of water which degrade the quality of the videos. As a contribution for enhancing underwater images, we develop an algorithm using a single denoising autoencoder to restore the color and remove the disturbances such as marine snow from underwater images. Marine snow in some images is only partially removed using the proposed network; however, we show the reason behind this failure. Related learning methods use generative adversarial networks (GANs) for generating color corrected underwater images, and to our knowledge this thesis is the first to deal with a single autoencoder capable of producing same or better results. Moreover, underwater aligned image pairs are established for the training of the proposed network, where it is hard to obtain such dataset from underwater scenery. The objective is to increase the accuracy and reliability on automatic underwater operations that rely on robotic perception without human interference by improving the quality of the captured frames. At the end, the proposed network is evaluated using Mean Squared Error (MSE), Peak Signal-to- Noise Ratio (PSNR), and Structural Similarity Index Metric (SSIM) quality metrics. Additionally, we compare our experiment with a related method. The proposed network takes into consideration the computation cost and the accuracy to have real-time implementation on visual-driven tasks using a single autoencoder.
Automated Object Pose Recognition by a Combination of Stereo Cameras and 2D Object Detection
Rostock, Univ., Master Thesis, 2018
The ability to visually recognize the 3D pose of objects would be helpful in many industrial applications. Existing approaches on this topic restrict the pose estimation to simplified scenarios, e.g. where a pose consists of just a 2D position and an angle around the vertical axis, or they require a priori knowledge. In this thesis methods are investigated to estimate a 6 DoF pose and 3D extends of arbitrary objects captured by a stereo camera. No knowledge about the shape of objects is required beforehand. A prototype implementation is provided. It first detects the objects in 2D employing a CNN trained for object detection. Then, the depth information is used to reconstruct a 3D point cloud and isolate the objects of interest. Finally, pose and size are estimated based on dense and sparse registration methods. At the end, the whole method is tested on artificially generated stereo images of fish. The results show remaining challenges especially regarding the robustness.
Resolving Classification Ambiguities in Convolutional Neural Networks Using Hierarchical Structures
Rostock, Univ., Master Thesis, 2018
We have recently witnessed the revolution of deep learning and convolutional neural networks enabled by the powerful machines available today. Convolutional neural networks have demonstrated excellent performance on various vision tasks, most importantly classification and detection. Nevertheless, there are some difficulties in the way of perfect performance. One problem is discriminating among objects that look extremely similar visually but semantically they are different. Another problem is the high cost of training large detection models. The same cost applies when the model is required to detect a new type of objects. In this work those problems are handled by introducing visual concepts and the use of hierarchical structures. We will see how the accuracy of classifying similar objects can be highly improved and how the time of accommodating for new objects in a detection model can be reduced from days to hours.
Nutzung von Stereo-Hochkontrast-Aufnahmen für eine bodengestützte, automatische Bestimmung von Schüttgutmaterialien und für die Erkennung der Belegung von Außenlagerflächen in Häfen
Go-3D 2017: Mit 3D Richtung Maritim 4.0
Go-3D <8, 2017, Rostock, Germany>
In diesem Artikel wird ein Stereokameraaufbau vorgeschlagen, welcher es ermöglicht, über weite Distanzen einen Schüttguthaufen freizustellen und seine Art (Sand, Kies, Roheisen) zu bestimmen. Durch Aufnahme von Belichtungsreihen werden Hochkontrastbilder gewonnen. Somit ist das System für Außenaufnahmen bei jedem Wetter geeignet. Der vorgestellte Ansatz arbeitet in fünf Schritten: Aufnahme von Stereo- Hochkontrastbildern, Bildvorverarbeitung, Semi-global Blockmatching, Segmentierung/Rücktransformation und Materialerkennung. Es wird in diesem Artikel gezeigt, dass dieses Verfahren robust gegenüber Wetterbedingungen ist, eine zuverlässige Segmentierung auf Basis der Tiefeninformationen erreicht und die korrekte Materialbestimmung ermöglicht.
Real-time Line Detection in Omnidirectional Images with an FPGA for Monitoring and Reporting Applications in Production Processes
Go-3D 2015: Computergraphik für die Praxis
Go-3D <7, 2015, Rostock, Germany>
In this paper we introduce an approach to FPGA-based line detection in omnidirectional images, taken by an upward-oriented camera. This approach can be used as a basic technology for indoor positioning and therefore for many applications which need to estimate the user positions inside electromagnetically shielded buildings or ship hulls. The detected lines can be compared with lines from the edge model which is derived from the existing CAD model of the surrounding room. The paper starts with a short summary of related work in the domain of indoor positioning, the mathematics behind line detection by the Hough Transformation (HT) as well as an explanation of the Omnidirectional Hough Transform (OHT). In the second part the context of the main contribution is described including the hardware setup and application requirements. After that a novel approach to line detection using the OHT on FPGA is presented and the OHT Module is introduced. Then we show the results and make a comparison between our FPGA implementation and the implementation on a desktop computer. The paper ends up with a conclusion and an outlook to future works.
Ereignisdetektion in Bildfolgen
Rostock, Univ., Diplomarbeit, 2014
Diese Diplomarbeit behandelt die Ereigniserkennung in Bildfolgen. Dabei soll der Anwendungsfall "Ereignisdetektion in an Bord von Fischereifahrzeugen aufgenommenen Überwachungsvideos" adressiert werden. Nach einem Überblick über die Grundlagen werden Konzeption und Umsetzung eines Systems erläutert, welches die Ereigniserkennung erleichtern soll. Bedingt durch den Anwendungsfall wird dabei vermehrt auf die Erkennung von Meerwasser eingegangen, welches einen wichtigen Kontext für viele Ereignisse da stellt.