If you want to sign up for this seminar, you need to register with the following information in the form

• surname, first name,
• e-mail address,
• matriculation number,

Please note that the registration period starts 03.10.2025 at 08:00 h and ends 24.10.2025 at 23:59 h. All applications before and after this registration period will not be taken into account.

Registration will be possible within the before mentioned time by sending an e-mail with the desired seminar topic, name and matriculation number to This email address is being protected from spambots. You need JavaScript enabled to view it..

Only one student per topic is permitted (first come - first serve).

The registration is binding. A deregistration is only possible by sending an e-mail with your name and matriculation number to This email address is being protected from spambots. You need JavaScript enabled to view it. until Sunday, 27.10.2024 at 23:59 h. All later cancellations of registration will be considered as having failed the seminar.

Students write a scientific report on a topic closely related to the current research of the DSS group.Therefore, potential topics include pattern recognition and machine learning related aspects.

Students will also present their findings in front of the other participants and the DSS group.

The scarcity of labelled data poses a significant challenge in developing effective deep learning models for classifying underwater acoustic signals. The shortage of large, publicly available datasets stems from the high cost of data collection, the confidentiality of acoustic recordings and the difficulty of accurately annotating data. Transfer learning is a technique in which knowledge from a model trained on a large source task (e.g. image recognition) is applied to a different, yet related, target task. It offers a promising approach to mitigating this issue. The aim of this seminar paper is to examine the transfer learning technique, comparing it with existing NN models to explore how it improves the performance of underwater radiated noise classification.

The assessment of speech quality and intelligibility is of great interest, both in speech therapy and in the evaluation of algorithms for improving speech signals with regard to interference factors such as noise and reverberation. The aim of this seminar paper is to summarize and compare various objective assessment methods. In addition to traditional assessment criteria such as STOI, PESQ, etc., approaches based on neural networks should also be used for the comparison.

Reinforcement learning (RL) can be applied to the control of autonomous SONAR systems. Since SONAR environments are typically noisy and partially observable, RL algorithms must meet special requirements regarding robustness and adaptability. There are different RL algorithms that are used for such control and decision-making tasks, including policy gradient methods (e.g., PPO), value-based methods (e.g., DQN), and combined approaches (e.g., SAC). These algorithms differ in terms of training and sample efficiency, stability, overall performance, and the types of control problems they are best suited to address. The aim of this seminar is to understand the individual algorithms, systematically compare them, and evaluate their suitability for SONAR-based applications.

Direction-of-arrival (DoA) estimation is a fundamental task in signal processing. Classical approaches such as beamforming are widely applied, for example with microphone arrays or hydrophones in underwater acoustics. In addition to these established methods, new sensor technologies are gaining increasing relevance. Of particular interest are optical fibers, which, through Distributed Acoustic Sensing (DAS), can detect and localize acoustic signals along the fiber. The aim of this seminar paper is to explore how beamforming methods can be implemented and enhanced using Machine Learning. A special focus will be placed on identifying synergies between the concepts of Distributed Acoustic Sensing and Machine-Learning-based beamforming approaches.

Passive SONAR systems continuously generate waterfall plots that display acoustic signals over time and frequency. Traditionally, their interpretation has relied on the expertise of SONAR operators, who must identify patterns such as propeller frequencies, machinery noise, or biological signals. With the integration of Artificial Intelligence, new opportunities arise to automate signal detection and classification, increasing both efficiency and accuracy. In this seminar your goal is to gain an overview of current research trends and future scenarios.

The Segment Anything Model is a powerful model for segmentation tasks. With its capability of temporal processing (since SAM 2) and its zero shot capability it could offer the ability to track objects in the water environment (whales, fishes, ships, etc.). The aim of this seminar work is to figure out if SAM 2 can be used in the underwater domain for segmentation in SONAR scans. If it is not directly applicable you should give some insights in possible extensions and solutions to this research question.

In a multistatic SONAR network, passive sonobuoys listen while an active source pings; targets are localized from time-difference-of-arrival (TDOA) measurements. In practice, harsh underwater conditions mean TDOA is often insufficient or noisy, so naïve geometric intersection quickly fails. Building on Shin et al., you will model a simplified 2D scenario (few receivers, limited SNR, missed/false detections) and study robust tracking when TDOA constraints are incomplete: compare baseline ellipse-/hyperbola-intersection and standard data association against likelihood-based track splitting and stack-based association, then explore how a small ML component (e.g., a classifier for association or a learned gating rule) can further stabilize tracks. Your goal is to propose and evaluate a pipeline that maintains accurate trajectories despite sparse, ambiguous TDOA.