Welcome to the repository for the AI-Driven Image Recognition Techniques workshop of the REDUCE project: Reducing Bycatch of Threatened Megafauna in the East Central Atlantic.
This three-day hands-on workshop, held from May 5 to 7, 2025, at the University of Barcelona (Spain), was designed to cover AI-driven image recognition techniques that can be used for marine species identification.
- Introduce core AI-driven image recognition techniques, from classic to state-of-the-art deep learning methods.
- Provide hands-on experience in implementing and fine-tuning deep-learning models for species identification.
- Discuss recent breakthroughs in Convolutional Neural Networks (CNNs) and Transformer-based vision architectures.
- Explore real-world applications in fisheries monitoring and bycatch mitigation.
Theme: Deep Learning Techniques for Image Recognition
Instructors:
• Catarina Silva, University of Coimbra (CFE), Portugal
• Ricardo Cardoso Pereira, University of Coimbra (CISUC), Portugal
| Session | Duration | Topics |
|---|---|---|
| Theoretical | 1 h | • Why deep learning excels in vision tasks • Neural-network components (layers, activations, etc.) • CNN building blocks: convolutions, pooling, fully-connected layers • Landmark architectures (e.g., ResNet) • Transfer learning & pre-trained models • Vision Transformers (ViTs) |
| Practical (Colab) | 2 h | Guided notebook: training & fine-tuning a CNN / ViT for species identification |
Theme: Evaluating Model Performance
Instructors:
• Nina del Rio Ares, Institute of Marine Research (IIM-CSIC), Spain
• Catarina Silva, University of Coimbra (CFE), Portugal
• Ricardo Cardoso Pereira, University of Coimbra (CISUC), Portugal
| Session | Duration | Topics |
|---|---|---|
| Theoretical | 1 h | • Performance metrics (accuracy, precision, confusion matrix, etc.) • Avoiding overfitting (dropout & data augmentation) • Cross-validation strategies (k-fold) |
| Practical (Colab) | 2 h | Guided notebook: model evaluation & error analysis |
You can run the practical component of the workshop both locally and on Google Colab.
In an environment running Python 3.11 and capable of opening Jupyter Notebooks, install the project dependencies using:
pip install -r requirements.txtAfterward, open the provided .ipynb file and run all cells.
Tip: A GPU-enabled environment (CUDA-capable) will speed up training, but is optional.
- Open Google Colab.
- Press File → Upload notebook and select the provided
.ipynbfile. - Upload the
imagesfolder and themodel_params.ptfile to the Colab "Files" pane. - Press Runtime → Run all.
Colab already includes the required libraries, so no extra installation is needed.
Tip: Selecting the T4 GPU-enabled runtime will speed up training, but is optional.
The images used in this workshop were obtained from GBIF - Global Biodiversity Information Facility.
GBIF is an international network and data infrastructure funded by the world's governments and aimed at providing anyone, anywhere, open access to data about all types of life on Earth.
Distributed under the GNU General Public License v3. See LICENSE for more information.