This repository is part of GLORiA Tools, a suite of utilities designed to support fish classification, segmentation, and analysis in aquaculture environments using deep learning techniques.
🔍 This module implements and evaluates Convolutional Neural Network (CNN) architectures for 3-class classification of fish images: wild, escaped, and farmed.
| Folder | Description |
|---|---|
App/ |
Graphical User Interface (GUI) for interactive fish classification. |
Cropping/ |
Automatic cropping tools based on color thresholds or segmentation masks. |
Models/ |
Pretrained CNN models and training scripts. |
Segmentation/ |
Scripts for fish segmentation and corresponding masks. |
Synthetic/ |
Synthetic data generation and advanced augmentation strategies for class balancing. |
preprocessing/ |
Tools for dataset cleaning, file renaming, and directory structuring. |
This module includes multiple CNN architectures trained to perform supervised classification across 3 fish categories.
It utilizes transfer learning, data augmentation, and robust evaluation metrics to benchmark performance.
- ✅ ResNet50
- ✅ VGG16
- ✅ MobileNetV2
This tool is integrated into M1: Benchmarking Deep Learning Models for Fish Classification, part of the GLORiA-Tools project. It is used for:
- Establishing CNN baselines to compare against CNNs, ViTs, and CLIP-based approaches for both binary and 3-class classification.
- Evaluating the impact of segmentation and preprocessing on classification accuracy.
- Comparing two-class setups (e.g., grouping escaped and farmed fish together) versus full three-class distinctions.