AI Research Intern - Generative Data Augmentation for Microscopy

The Internship In Short

Internship topic: Bridging the lab-to-field gap in microscopic spore detection through generative data augmentation.

Join the artificial intelligence team at Multitel and explore the latest advances in domain adaptation, 
image-to-image translation and synthetic data generation for microscopy. During this internship, you 
will build and validate a synthetic data pipeline that makes lab-grown fungal spores look like spores 
captured in the field, so detection models trained on synthetic data work on real field images. You will 
work with state-of-the-art generative and computer vision techniques and contribute to research on AI 
for sustainable agriculture.

The topic

Trans-e-Bio, part of the Biocontrol 4.0 initiative, develops digital tools to support the adoption of biocontrol solutions in agriculture across the France-Wallonia-Flanders border region. By combining local risk detection with predictive models, the project helps farmers time their treatments better and reduce synthetic pesticide use. A key building block is the automatic detection of Venturia inaequalis ascospores, the cause of apple scab, in microscope images collected in the field.

Annotated field images are scarce, so one solution is to generate synthetic training data by compositing lab-grown spores onto real field backgrounds. Previous work at Multitel showed that lab and field spores have very similar shapes but very different appearance: shape descriptors overlap strongly, while 33 of 40 texture descriptors differ significantly. Detection models trained on these composites transferred poorly to real field images. Whether this comes from the appearance gap, the compositing procedure or the detection architecture has not yet been established.

To close this gap, two families of approaches can be considered. Learned approaches, such as unpaired image-to-image translation or diffusion-based appearance transfer, learn the appearance of field spores directly from data. Physically motivated degradation models, such as histogram matching, point spread function (PSF) estimation and noise modelling, reproduce the imaging conditions of the field. Both will be compared.

What you will work on:

You will explore several research questions, including:

• Where does the drop in performance from synthetic to real field images come from: appearance, compositing, or model architecture?
• Can generative models or physically motivated degradations make lab spores look like field spores?
• How can the remaining domain gap between synthetic and real images be measured?

As an intern, you will play an integral role in the research and development of this synthetic data pipeline. Responsibilities include:

• Conducting an in-depth literature review on domain adaptation, image-to-image translation and photometric augmentation for microscopy;
• Reproducing the existing pipeline and establishing a validated detection baseline on real field images;
• Selecting, implementing and integrating translation and augmentation methods into the compositing pipeline;
• Training instance segmentation models (YOLACT++ and comparable architectures) on the resulting composites;
• Evaluating performance on held-out, expert-annotated field images, including measures of the remaining domain gap;
• Running ablation studies and improving the methods based on the results;
• Documenting and presenting your results.

The profile we are looking for:

We are looking for a motivated student who is curious about AI research and enjoys turning ideas into working solutions.

You should have:

A background in Computer Science, Artificial Intelligence, Data Science, or a related field, typically through an engineering degree or university program.

• A strong interest in Computer Vision and Deep Learning, particularly in recent developments in generative models and domain adaptation.
• Skills in signal and image processing.
• Fluency in English is mandatory; proficiency in French is a plus.
• Experience with machine learning or AI projects.
• Good programming skills in Python.
• A basic foundation in probability and statistics.
• Knowledge of PyTorch or another deep learning framework.
• Familiarity with OpenCV or other image processing libraries is a plus.
• Basic knowledge of Git and software version control.
• Working proficiency with Linux and Docker is a plus.
• A self-starter mindset, with the ability to work both independently and as part of a team.
• A curious and analytical approach to solving complex problems, with good synthesis skills.
• Attention to detail and commitment to producing robust, reproducible results.

Experience with GANs, diffusion models or instance segmentation is a plus, but a strong motivation to learn is equally valued.

Even if you don’t check every box in our requirements, we encourage you to apply. We value diverse backgrounds, curiosity, and the drive to learn, grow, and achieve.

Recruitment Process:

1. Application Review

We review your CV and cover letter to assess your background, skills, and interest in the internship topic.

2. Interview

Selected candidates are invited to an interview combining a technical discussion with a conversation about the internship objectives, your interests, and your expectations.

3. Following the interview

We select the candidate whose profile and interests best match the internship.