Master's Thesis Intern Sensor-Agnostic Raw Image Denoising
08/2026 – present In collaboration with the Chair of Media Technology, TUM
- Developing a raw image denoiser that generalizes across camera sensors without per-sensor retraining
- Modeling and synthesizing sensor noise to generate realistic training data
- Designing the network for efficient deployment on embedded hardware
Research Intern (Forschungspraxis) Joint Image Denoising and Compression
03/2026 – 04/2026 - Investigated single-frame denoising directly in the latent space of a learned image codec (pretrained convolutional autoencoder, frozen encoder/decoder), combining denoising and compression in one pipeline
- Showed that latent-domain denoising outperforms the equivalent pixel-space baseline and that latent quantization adds no denoising loss, so both fit one pipeline at no extra bitrate
- Identified model size as the barrier to edge deployment and proposed a multi-frame extension that exploits temporal redundancy to reach the same quality with smaller models
Working Student Edge AI Deployment & Video Pipelines
07/2024 – 07/2026 - Deployed real-time computer vision models (YOLO detection and segmentation) on the NXP i.MX 95 (Arm CPU + eIQ Neutron NPU) with TensorFlow Lite and the Neutron delegate; built the full C++ inference path (preprocessing → inference → postprocessing) at ≥ 30 FPS
- Designed and implemented GStreamer video pipelines for streaming, recording and AI inference running concurrently
- Engineered buffer-level flow control, including temporal filtering across frames and display-buffer management, to keep the real-time pipeline smooth under load
- Built real-time rendering with interactive UI overlays, driven by asynchronous event handling and a multithreaded architecture
Bachelor's Thesis Intern DNN Efficiency in NPU-Accelerated Image Processing
04/2024 – 06/2024 - Optimized CNN-based denoising models for NPU deployment using neural architecture search (NAS), pruning and quantization, and profiled their performance and efficiency on the target AI accelerator
- Matched the image quality of classical denoising methods while running within the target hardware's compute budget
Working Student AI-Driven Image Signal Processing
01/2024 – 03/2024 - Tested and ran inference of deep learning models for AI-driven ISP, such as denoising and end-to-end ISP
Working Student Mixed-Signal IP
12/2022 – 12/2023 - Measurement setups, soldering and bare-metal programming of test boards