Starting date: September 2026
Application deadline: Open until the position is filled
Duration: 36 months
Net salary: Approximately €1810 per month.
Workplace: Lab-STICC UMR 6285 CNRS, Brittany INP – ENIB, Brest, France
Funding: Fully funded by Brittany INP through a French ministerial doctoral contract
Context of the PhD position:
Lab-STICC is a research unit affiliated with the French National Centre for Scientific Research (CNRS). Its staff members, more than 650 people, are distributed across several institutes and geographical sites in Brittany. The unit conducts research around the central theme « from sensors to knowledge ».
We are seeking an outstanding PhD candidate with a strong background in computer vision and machine learning to work in close collaboration with Ifremer on the design of fast, robust, and interpretable algorithms for automatic fish species detection and classification from underwater imagery.
Understanding the impact of human activities and climate change on exploited marine ecosystems requires a deeper knowledge of marine biodiversity and ecosystem dynamics under natural conditions. Establishing reference catalogues is therefore essential for monitoring future environmental changes, assessing biodiversity loss, and supporting evidence-based conservation strategies. For example, evaluating fishing pressure on a given fish species requires reliable information about stock levels and the capacity of that stock to renew itself through reproduction. Reference data make it possible to determine whether fish populations are in a healthy or degraded ecological state. When resources are overexploited, overfishing can alter species abundance and diversity, disrupting the balance of marine ecosystems. Studying and better characterizing the seabed is therefore not only a scientific necessity but also a key step toward protecting marine biodiversity and ensuring the sustainable management of marine resources within a responsible and environmentally respectful economy.
Objectives and challenges:
Automatic underwater object recognition can be divided into two main stages: object detection, which aims to localize each object in an underwater image, and object classification, which aims to identify the class or category of each detected object. These tasks remain a challenging problem in pattern recognition because underwater images are often degraded by light absorption, scattering, low contrast, color distortion, turbidity, blur, and motion-induced artifacts.
In addition, illumination can change rapidly because of water-column dynamics, visibility is often limited, and complex backgrounds may vary due to moving aquatic plants or suspended particles. Fish can move freely in all directions, appear under different viewpoints, overlap with one another, or hide behind rocks and algae. Similarities in body shape, texture, and color patterns across different species further complicate fish species recognition.
Previously, we developed supervised deep learning algorithms for underwater fish detection and fish species classification. Despite their promising results, fully supervised approaches depend heavily on large annotated underwater datasets, whose acquisition and labeling are costly and time-consuming. The scarcity of large-scale, high-quality annotated datasets also limits the generalization ability of supervised models.
To address these limitations, our recent work has increasingly explored semi-supervised learning, which leverages both labeled and unlabeled data to improve model performance while reducing reliance on manual annotation. Within this PhD thesis, we will investigate self-supervised learning and vision foundation models to exploit large-scale unlabeled underwater data and develop more accurate, robust, and generalizable fish recognition systems.
This PhD thesis is structured around three research questions addressing the development, interpretation, and efficient deployment of self-supervised deep learning models for underwater fish detection and classification.
Q1. How can we develop robust and transferable self-supervised models for underwater fish recognition?
Q2. How can we make AI-based fish identification explainable for biologists and taxonomy experts?
Q3. How can we improve the efficiency of the developed models and reduce their computational and environmental footprint?
Expected impact:
This work will advance the understanding and practical use of deep neural networks for underwater object detection and classification. It will provide more robust models for fish species recognition, more transparent decision-making mechanisms for expert interpretation, and more efficient solutions for real-world deployment. The findings are expected to contribute to marine biodiversity monitoring, ecological studies, and environmental decision-making by enabling faster, more reliable, and more sustainable analysis of underwater visual data.
Required qualifications and skills:
- Master’s degree in computer science or related field with specialization in computer vision/machine learning.
- Experience within deep learning frameworks is highly recommended.
- Knowledge of programming languages: Python, C++.
- Knowledge of libraries: PyTorch, TensorFlow, OpenCV, PCL.
- Proficient in English language (written and oral).
- Interpersonal skills and the ability to work in a multidisciplinary team are recommended.
- Taste for research activities in submarine applications is a plus.
How to apply:
Interested candidates should send the following documents in a single PDF file to abdesslam.benzinou@enib.fr ikram.kourbane@enib.fr and kamal.nasreddine@enib.fr :
- Curriculum vitae.
- List of courses and grades.
- Cover letter explaining the candidate’s interest in the position and how their skills fit the doctoral project.
- Copies of diplomas.
- Names and contact details of at least two reference persons.
