In this paper, we present our comprehensive approach to the implementation of extracurricular STEM courses, drawing from 17 years of experience in the field. Our methodology encompasses various aspects, including administrative procedures, strategies for promotion, and didactic methods. Additionally, we provide insights into participant numbers and a practical illustration of our didactic approach through specific course examples, offering a nuanced understanding of our proposed framework.

In this paper, we propose a novel method for vectorizing Wave Digital structures using Automatic Differentiating Wave Digital Filters (ADWDFs), maintaining modularity through a backend/frontend construction approach, and enabling parallelization as well as hardware acceleration by design. As demonstrated by a proof-of-concept example, our approach achieves up to a 10 times speedup compared to a non-vectorized reference implementation.

In existing semi-supervised learning (SSL) approaches, the contributions of labeled and unlabeled data points to the training objective are expressed as two separated terms for the respective sets. A single weight specifies how much the unlabeled samples contribute to the overall loss and is carefully scaled in the course of the training. In this work, we propose a SSL framework in which the contribution of each unlabeled sample is individually scaled based on the confidence generated by a teacher model. We demonstrate that our scaling mechanism enables the student model to efficiently learn from heavily interpolated and augmented samples and achieve state-of-the-art performance.

Due to their ability to model the manifold of natural images, generative adversarial networks (GANs) have been adapted to semi-supervised learning and shown promising results. Using this property of GANs, manifold regularization has been incorporated into feature-matching GANs (FM-GAN) [14] and shown improvement over FM-GAN. In this work, we present a novel approach for performing manifold regularization in semi- supervised classification using the FM-GAN framework. Our approach, mixup manifold regularization (MUMR), regularizes the networks to favor linear behavior so that linear interpolations of noise vectors lead to linear interpolations of the associated predictions. We show that our approach is able to achieve state-of-the-art results when compared to other GANs- and perturbation-based approaches.

With an increasing significance of climate change and the progressive development of new intelligent devices in the sense of an Internet of Things (IoT), attempts are being made to recover unused ambient energy and convert it into electricity in order to supply said devices, known as energy harvesting. To raise awareness to the topic among students and next-generation engineers, we propose an energy harvesting based teaching platform for physical computing that allows students to experience the concept and design of energy harvesting systems playfully and creatively. Therefore, a self-powered energy harvesting hardware (gaming) platform (EHHP) has been developed involving a variety of components and peripherals which enables a wide range of software and hardware development teaching possibilities in physical computing. Piezoelectric elements and amorphous silicon solar cells in combination with supercapacitors were selected as energy sources which sufficiently enable several hours of operation under gameplay conditions. Besides, a software framework is presented that can be adapted to the level of experience or age of the students and can thus be kept completely modular. Further, since the concept involves mechanical design, CAD construction in combination with rapid prototyping by means of 3D printing techniques can be conducted by students.

The applications of Machine Learning and Neural Networks (NN) are nearly unlimited and the application of artificial intelligence has gained much interest in recent years, because of their great performance in various tasks. Deep Convolutional Neural Networks (CNNs) are a state-of-the-art technique for visual recognition in image- and video data. However, the application range of a specific CNN is very limited, because the CNN is adapted for a specific task with an exclusive dataset for training. It needs to be rebuilt from scratch when the input- or output parameters are just slightly changing, including the collection of a new dataset for training. To reduce those cost and time expensive issues, transfer learning can be beneficial, where the outcome of an already pre-trained Neural Network, the source data, is fitted to the target dataset of a new task. In the case of object recognition, there are several use cases where pre-trained Deep CNNs are applied. But those Deep CNNs can not only be used for visual recognition. In this work the approach is made, to use transfer learning on DCNNs for spoken letter recognition, although the target data is very dissimilar from the source data, to show the range of application for transfer learning. Moreover, this application is trained with a very small dataset.

Implementation and evaluation of image/video processing systems on FPGA hardware require a video path for input/output of data. We propose a portable low-cost video pipelining method by means of a USB 2.0 interface in conjunction with a Raspberry Pi single board computer. An application of our approach is demonstrated by a 2-D spatio-temporal Wave Digital Filter for motion feature extraction.

Implementation and evaluation of image/video processing systems on FPGA hardware require a video path for input/output of data. We propose a portable low-cost video pipelining method by means of a USB 2.0 interface in conjunction with a Raspberry Pi single board computer, whereby virtually any video sources/sinks can be used as input/output with realtime capabilities. An application of our approach is presented by the implementation and evaluation of a 2-D spatio-temporal Wave Digital Filter for motion feature extraction.

In recent years the research community and industry have paid high attention to the field of machine learning, especially deep learning. Nowadays many real-world classification or rather prediction applications are implemented by neural network models. We propose a multi-view fusion neural network with application in the manufacturing industry. Image information of multiple cameras is fused and used by the proposed model to predict the state of a manufacturing machine. Experiments show that the overall classification performance is increased from a baseline of 92.7% to 99.5% by the fusion model.

This paper presents the combination of the adversarial nets framework with a kernel-based statistical dependency measurement for learning interpretable representations without any labeled data and with an autoencoder. Thus, the loss function consists of three terms and, based on the loss function, a new architecture emerges. Experiments show the functionality and potential of the architecture and the proposed kernel-based statistical dependency criterion quantitatively and qualitatively. Furthermore, the improvements are presented on a dataset consisting of industrial images.

Implementations of wave digital filters (WDF) require quantization of finite word lengths. We present a method for finite word length optimization of a 2-D-WDF, using magnitude truncation for quantization and weighted quantization error variances, constrained by a limited shared memory bus width. Moreover, we define a rule to obtain word lengths for various image sizes and arbitrary bus widths, whereby long simulation times of the filter can be avoided during design time. After all, an application of our approach offers a constant improvement by 23 dB, compared to an intuitive unbalanced approach.

Many n-D signal processing applications in industrial environments require real-time processing of data. We propose a technique for realization of motion feature extraction in video images on low performance computing hardware. The realization is based on a 2-D spatio-temporal wave digital filter (WDF) using SDRAM-based shift operators. This enables the usage of different effective shared bus widths for the connection between a DDR2-SDRAM and the WDF arithmetic subsystems, which is demonstrated by two different configurations. The FPGA board is processing real-time video signals from an image sensor and the filtered output can be observed on a display.

Increasing miniaturization of electronic systems and nano-technology enables more and more the realization of real time 3-D spatio-temporal signal processing and control systems. However, while stability is well understood for 1-D systems, open problems remain in the higher dimensional case, e.g. for 3D system realizations, especially with respect to computational complexity and thus hardware effort. We propose a basic and thus fast stability test for systems that can be given in a standard 3-D Roesser-like state space model. Most practical systems can be represented in that manner or at least be transformed into an equivalent system so that the test is applicable to most real world problems. The test itself is inspired by passivity assumptions which for example guarantee stability of electrical reactance networks. It represents a sufficient condition for stability and comprises eigenvalue computations of a sum of Gramian matrices of the same size as the system matrix, which leads in this step to the same complexity as a 1-D stability test. The latter can thus also be applied to adaptive systems, where the system matrix is periodically changing.

Visual tomography is a relatively new method for 3D scene reconstruction. It is adopted from medical tomography and based on multiple images from different viewpoints of a scene. In this context, multidimensional spectra and filtering techniques are the key technology for the reconstruction process. Visual tomography differs from classical tomography in several aspects which leads to new challenges with respect to mathematical description. The present paper examines the influence of image background on reconstruction quality. This background problem does not appear in classical medical tomography applications. In particular, the influence of multidimensional sampling and restrictions with respect to the number of view angles can be analyzed by using multidimensional signal theoretical concepts. The differences between ideal (no background) and real acquisition conditions are examined. Visual tomography has the potential for innovative new fields of applications, where Enabling Technologies for Societal Challenges are the focus of our considerations. Demographic change leads to a high interest for enabling mobility for elderly people with physical disabilities. Walking frames equipped with such technologies will be able to assist such people in real day environments.

This paper presents a new hardware-efficient method of eliminating aliasing artifacts in time direction occurring in filter results of velocity or directional filters applied to image sequences recorded by a camera. Aim is to deliver reliable input data for object detection in a driver assistance system, for example. Applications designed for mass products like these demand low production costs and therefore cheap sensors and hardware. Thus, resolution as well as frame rate of the captured image sequences are often limited and therefore Shannon's sampling theorem is not always satisfied. For the case that aliasing only occurs in the pass band, a method to remove unwanted aliasing components from the filter result without any additional filter hardware is proposed and demonstrated by means of examples.

Stability of multidimensional (k-D) systems is still a challenging field of work. Well known and established stability measures may lead to complex mathematical problems, while simple tests are restricted to special cases of n-D systems. A new stability test for certain discrete 2-D system realizations given in a Roesser model description is proposed. This test is suitable for signals bounded with respect to both coordinate directions, like images. The 2-D system is observed in real operation, i.e. considering a sequence of processing, which leads to a 1-D state space description. The resulting 1-D system matrix is some kind of a block Toeplitz matrix that allows definition of the test.

In this paper a new anti-aliasing approach for motion-based object detection using linear shift invariant (LSI) filters is presented. Mainly originating from the field of signal processing, LSI-filter-based motion detection has been topic of research for a long time, though due to low computational power of contemporary computers the developed systems have been unfeasible for application in mass products like driver assistance systems. However, recent progress in hardware in conjunction with decreasing costs makes using linear filters a very interesting alternative of increasing performance of common approaches. One of the most important factors falsifying object detection results obtained using linear shift invariant (LSI) is aliasing. This is mostly caused by the low temporal resolution of the camera signal which leads to the fact that the sampling theorem does not hold for objects moving at high speed in the image plane. In general, aliasing errors cannot be removed after sampling. Under certain conditions, however, their impact on the filter result can be successfully decreased using the approach presented in this paper. The applicability of the new approach is demonstrated using scenes recorded by a camera installed in a blind spot warning system.

In practice, n-D signal processing problems often show the property that underlying signals are unbounded with respect to only one direction, like the spatio-temporal signal of a line-scan camera, for example. Thus, the use of systems which meet the conditions of practical BIBO stability, mainly known from the field of control theory, seems to be feasible. In this paper the practical BIBO stability concept introduced by Agathoklis and Bruton [1] is analyzed with respect to its applicability in the field of signal processing applications. It is shown that practical stability in its original form is not sufficient in signal processing if not further conditions are supposed. This is done by comparing the 2D Fourier transform of the actually measured impulse response of exemplary systems to the frequency response expected on basis of the transfer function.

Vision-based object detection is a core part of many automotive collision warning systems. Especially due to the extensive information an image can hold and the decreasing costs of computational power, it has attracted more and more interest in the last decade. In this paper we propose a motion-based approach to simplify the detection of moving objects in order to improve available methods and make them more efficient. We interpret the image sequence containing the moving object (e.g. a vehicle or a crossing pedestrian) as a three-dimensional signal, not just as a sequence of image matrices. That way we can benefit from the advanced theoretical knowledge on system description and handling from the field of signal processing. In detail, three-dimensional velocity filters implemented using wave digital filters are used to oppress every object in the image that is not moving in a certain direction at a certain velocity.

Due to the increasing calculation power of dedicated hardware, real-time image processing becomes practicable at reasonable expense. Nevertheless the algorithms concerned have to be adapted to real-time requirements. The subject treated in this paper is tie detection for automated railway inspection. An edge detector is proposed in which the sensitivity is adaptable to the local brightness of the processed image. The required effort is the performance of a convolution operation with subsequent decision concerning the sign of the values obtained.