@misc{Afzal2025, type = {Master Thesis}, author = {Afzal, Umair}, title = {AI-Driven Comic Generation: Exploring the Creative Potential of Generative AI in Digital Storytelling}, url = {http://nbn-resolving.de/urn:nbn:de:hbz:1383-opus4-22581}, school = {Hochschule Rhein-Waal}, pages = {111}, year = {2025}, abstract = {Creative sectors have accepted AI into their industry for establishing modern methods of generating digital content. This study introduces a novel AI-based framework based on AI technology, using GANs along with Stable Diffusion models to automate comic development. This study looks at juxtaposing narrative outlining through text with automated visual generation toward an integrated system that produces adaptable comic panels with appropriate visual structure. The research methodology that the project has followed can be built on three cornerstones: advanced GAN schemes for text generation and pre-processing, followed by image synthesis through Stable Diffusion. A specially developed algorithm for speech bubbles determined the optimal placement of that text, so it would function well and maintain a semblance of beauty. By iteratively refining and tuning the model, this system was evaluated. Initial observations regarding visual coherence and narrative alignment were hopeful, but further tests using quantitative metrics-for instance, FID for images and BLEU for text, as well as broader user feedback-would be needed to validate the efficacy of the model entirely. The impediments were, however, GAN mode collapse, irregular speech-bubble layout, and inconsistent artistic styles. Further research would uncover the potential role of AI systems in easing the comic generation process for creators, educators, and designers of digital content to enhance accessibility and efficiency. This method shows promising applicability in various domains like automated narratives, customizable comics, and educational material. Further along the way, the scientists plan to improve panel storytelling, create an intuitive interface, and expand the dataset to include more artistic styles. Such enhancements could maximize the gain from automated comic conception.}, language = {en} } @masterthesis{Sueyruege2026, type = {Bachelor Thesis}, author = {S{\"u}yr{\"u}ge, Ey{\"u}p}, title = {Vergleich von FPGA und Mikrocontroller f{\"u}r die Echtzeitverarbeitung von Audiosignalen zur Richtungserkennung mittels FFT und TDOA einschließlich LED-basierter Visualisierung}, url = {http://nbn-resolving.de/urn:nbn:de:hbz:1383-opus4-23229}, school = {Hochschule Rhein-Waal}, pages = {57}, year = {2026}, abstract = {Das Ziel dieser Arbeit ist es, ein praxisnahes System zur Echtzeitverarbeitung von Audiosignalen mit dem Fokus auf der Richtungsbestimmung von Schallquellen zu entwickeln. Als Hardwareplattformen werden daf{\"u}r ein Field Programmable Gate Array (FPGA) und ein Mikrocontroller (MCU) eingesetzt. Dabei konzentriert sich die Arbeit auf die Untersuchung der beiden Plattformen, wobei die Leistungsf{\"a}higkeit, Genauigkeit und Ressourceneffizienz verglichen werden. Zu diesem Zweck werden Algorithmen der Signalverarbeitung, besonders die Fast Fourier Transform (FFT) und die Time Difference of Arrival (TDOA), implementiert und ausgewertet.}, language = {de} } @masterthesis{Muesch2025, type = {Bachelor Thesis}, author = {M{\"u}sch, Yannick}, title = {Kolmogorov-Arnold-Transformer: Untersuchung hybrider KAN-Transformer-Architekturen im Hinblick auf Skalierung, Effizienz und Interpretierbarkeit}, url = {http://nbn-resolving.de/urn:nbn:de:hbz:1383-opus4-23376}, school = {Hochschule Rhein-Waal}, pages = {85}, year = {2025}, abstract = {Transformers have become the de-facto standard architecture in Machine Learning, particularly in Natural Language Processing. However, they are computationally expensive to train, with approximately two-thirds of non-embedding parameters residing in Multi-Layer Perceptrons (MLPs). Liu et al. (2024) revisited the long-disregarded Kolmogorov-Arnold representation theorem and demonstrated promising results in specific domains such as symbolic regression and PDE solving. Whether Kolmogorov-Arnold Networks (KANs) generalize to other domains, particularly NLP, remains an open research question. This work systematically evaluates five architecture configurations: MLP baseline, MLP with B-Spline activation, KAN with B-Spline, KAN with Mean aggregation, and Group-Rational KAN (GR-KAN) - across three model sizes (15M, 41M, 124M parameters). Experiments are conducted on text classification ("AG News") and language modeling ("FineWeb") tasks. Identical components (attention, embeddings) isolate performance differences to the feed-forward topology. Statistical validity is ensured through multiple random seeds and parameter-matched comparisons. KAN-based architectures achieve performance parity with MLPs on classification tasks (±0.5 percentage points accuracy). However, they consistently underperform on language modeling, with perplexity increases of +7 to +28 points. Training efficiency overhead ranges from 1.5-2.0× for KANs, while GR-KAN approaches baseline speed. Notably, KANs exhibit significantly higher sparsity (≈ 90\% vs. ≈ 40\%), enabling compression factors of 1.4-1.6× with minimal performance degradation. The MLP+B-Spline control condition reveals that performance deficits stem from KAN topology rather than activation functions. KANs present a viable alternative for discriminative NLP tasks but do not outperform MLPs for generative language modeling. The postulated interpretability advantage through learnable activation functions could not be empirically confirmed - learned functions degenerate to quasi-linear transformations. Future research should explore alternative basis functions and selective hybridization strategies.}, language = {de} }