Imaging · Vision · Machine Learning

Jasser Kraiem

I take imaging problems from the mathematical model to the deployed system: I first lay the theoretical foundations, then turn them into a solution that fits the hardware budget and does what the customer needs. I'm a master's student in Electrical Engineering at TUM, writing my thesis on sensor-agnostic raw image denoising at NXP Semiconductors with the Chair of Media Technology.

Graduating 2027 · looking for a PhD, including industrial PhDs

View CV Email LinkedIn

Portrait of Jasser Kraiem

NXP Semiconductors · master's thesis · 2026 – now

Sensor-Agnostic Raw Image Denoising

The foundation
Denoising RAW data, before the ISP pipeline, can restore far more of the original signal than working on processed RGB. But RAW noise is device-specific, so deep models stay biased toward the sensors they were trained on and degrade on unseen cameras.
In practice In progress
A camera-agnostic, controllable RAW denoiser that generalizes across sensor noise characteristics while staying efficient enough to deploy on embedded hardware.

NXP Semiconductors · working student · 2024 – 2026

Real-Time Vision Pipelines

The foundation
A model only matters once it runs in a live video stream: every frame has to be captured, preprocessed, inferred, postprocessed and displayed within its time budget, without the stages stalling each other.
In practice
Deployed YOLO detection and segmentation in C++ on the i.MX 95 NPU with TensorFlow Lite at 30+ FPS, inside GStreamer pipelines that stream, record and run inference, with buffer-level flow control and multithreaded rendering.

TUM Data Innovation Lab · industry project with MARTIN GmbH · 2025 – 2026

Thermal Camera Analysis for Waste-to-Energy Plants

The foundation
Starts from the physics: fresh waste is a cold mass inside a hot furnace. Detects and sizes it with per-cycle minimum tracking, Otsu thresholding and a thermal mass index.
In practice
Turns raw, noisy infrared footage, full of flames, ash clouds and moving grates, into waste-drop signals for plant operation, with ash-obscured frames detected and filtered first.

Chair of Media Technology, TUM · seminar paper · 2026

End-to-End Learned Image Compression: Approaches, Deployment, and Standardization

The foundation
Starts from the model: the VAE-based framework most learned codecs build on, and the advances in transforms, quantization and entropy modeling that extend it.
In practice
Follows the chain to the end: benchmarks representative models on a Tesla T4 GPU to judge real deployability, then examines JPEG AI, the standard that carries learned compression from research into products.

NXP Semiconductors · bachelor's thesis · 2024

NPU-Efficient Image Denoising

The foundation
CNN-based denoising, compressed with neural architecture search, pruning and quantization.
In practice
Matched the image quality of classical denoising methods while running within the target NPU's compute budget.
  1. 2024 – 2027 M.Sc. Electrical Engineering and Information Technology Technical University of Munich
  2. 2022 – now Working student, then master's thesis NXP Semiconductors · AI-driven image signal processing, edge AI video pipelines, raw denoising
  3. 2021 – 2024 B.Sc. Electrical Engineering and Information Technology Technical University of Munich · thesis on DNN efficiency for NPU-accelerated image processing
  4. 2020 Baccalaureate in Technical Sciences, ranked 1st nationwide Pioneer High School of Monastir, Tunisia

Full CV →

In progress Master's thesis, NXP Semiconductors & Chair of Media Technology, Technical University of Munich · 2026–2027

Sensor-Agnostic Raw Image Denoising

Raw image denoisers are usually trained for one camera sensor and have to be retrained for the next. This thesis develops a denoiser that generalizes across sensors without per-sensor retraining, by modeling and synthesizing sensor noise to generate realistic training data, and designs the network for efficient deployment on embedded hardware.

Inference pipeline of the VAE-based codec: analysis transform, quantization, entropy coding with a learned entropy model, and synthesis transform.

Seminar paper, Chair of Media Technology, Technical University of Munich · 2026

End-to-End Learned Image Compression: Approaches, Deployment, and Standardization

Jasser Kraiem, Zongxie Chen

A review of learned image compression (LIC): the VAE-based framework most models build on, advances in transforms, quantization and entropy modeling, and perceptual compression beyond rate-distortion. Representative models are benchmarked on a Tesla T4 GPU to assess practical deployability, and the JPEG AI standard is examined. Standardization is promising, but future work has to balance compression performance against deployability, for instance through adaptive compression.

The theory side of my work. From Probabilistic Modeling to Generative Modeling: six posts that build from probability basics to diffusion models, deriving each model from first principles. Best read in order. If you only have time for one, read Diffusion Models Explained: it shows how a model generates images by learning to remove noise, step by step.

  1. Part 02 Gaussian Mixture Models Explained Gaussian Mixture Models express a probability distribution as a weighted combination of Gaussians, capturing multi-modal data through interpretable components. 8 Mar 2026 · 12 min read
  2. Part 03 Variational Autoencoders Explained A Variational Autoencoder (VAE) is a generative model that learns a compressed, continuous representation (a latent space) of data. It consists of an encoder network that maps data to a distribution in the latent space and a decoder network that reconstructs data from samples drawn from that latent distribution. 12 Mar 2026 · 11 min read
  3. Part 04 Normalizing Flows Explained Normalizing Flows turn a simple Gaussian into a complex distribution through invertible transformations, enabling exact and tractable likelihood computation. 15 Mar 2026 · 8 min read
  4. Part 05 Generative Adversarial Networks Explained How GANs learn by pitting a generator against a discriminator: the minimax game, and why adversarial training replaced explicit density estimation. 19 Mar 2026 · 9 min read
  5. Part 06 Diffusion Models Explained How diffusion models generate images by reversing a gradual noise process: forward and reverse Markov chains, DDPM, and why they rival GANs on sample quality. 21 Mar 2026 · 12 min read