Fish Species Classification from Photos using CNN

Fish markets, landing centres and fisheries departments identify species by eye — fast for experts, inconsistent for everyone else, and a bottleneck for catch documentation. This project trains an EfficientNet-B0 classifier on about 2,200 fish photographs across six Indian freshwater species, with the fish region isolated before classification so market backgrounds don't become a shortcut feature. Students learn fine-grained visual classification, why background bias is dangerous in market photography, and how to evaluate species pairs that look alike (Rohu vs Catla juveniles). The demo classifies sample fish photos with representative predictions and species cards.

1. Fish photographs are collected from markets and landing centres and labelled by species; clear side views are preferred.
2. The fish region is isolated from the background with segmentation before classification.
3. Crops are resized to 224×224, normalized and split into stratified train/validation sets.
4. EfficientNet-B0, pre-trained on ImageNet, is fine-tuned with a new classification head and dropout.
5. Training uses Adam with early stopping; the held-out split gives accuracy and per-species metrics.
6. The web demo classifies sample fish photos through the same pipeline with representative predictions.
7. Species cards with scientific names and identifying features accompany predictions.

Tech stack:

• Python 3, TensorFlow/Keras
• EfficientNet-B0 (transfer learning)
• OpenCV (segmentation)
• NumPy, Matplotlib, scikit-learn
• Jupyter Notebook (training)
• HTML/CSS/JS classification demo
• Pillow

Dataset & model details

• Dataset: Custom curated freshwater-fish set — ~2,200 fish photographs from markets and landing centres in West Bengal and Maharashtra, labelled into 6 species: Rohu, Catla, Mrigal, Hilsa, Tilapia, Common carp; clear side views preferred; stratified 80/20 split.
• Task: 6-class fine-grained image classification; input = 224×224×3 fish crop, output = species probabilities.
• Model: EfficientNet-B0 backbone (ImageNet pre-trained) → GlobalAveragePooling → Dropout(0.3) → Dense(6, softmax); ~5.3M parameters; segmentation upstream.
• Metrics: Species accuracy ~91% (design target), top-2 accuracy ~96% (design target), per-species precision/recall. No accuracy is claimed as measured until the training run is executed for the order.

• Six-species classifier Rohu, Catla, Mrigal, Hilsa, Tilapia, Common carp — the commercially important Indian freshwater species, with per-species metrics.
• Fish-region isolation The fish is segmented from market backgrounds before classification, removing a major source of dataset bias.
• Curated species dataset About 2,200 fish photographs from markets and landing centres, labelled by species with clear side views preferred.
• Species cards Each prediction shows the scientific name and key identifying features — genuinely useful for learning the species.
• Interactive classification demo Single-file web app: pick a sample fish photo and see the predicted species with probability bars.
• Lookalike-pair evaluation Confusion matrix documents Rohu/Catla confusion honestly — the hard case every examiner asks about.
• Viva kit Report PDF, PPT and Q&A covering fine-grained classification, segmentation preprocessing and evaluation.

• Curated, labelled fish-species dataset with split files
• Segmentation + training Jupyter notebooks
• Trained EfficientNet-B0 model with preprocessing code
• Interactive classification web demo with sample photos
• Confusion matrix and per-species metrics
• Species reference cards
• Project report PDF, PPT presentation
• Viva Q&A preparation document

• Juvenile fish and damaged specimens are harder; the report quantifies this drop.
• Clear side views are recommended — top-down or occluded fish degrade accuracy.
• 91% is a design target, stated honestly — the report documents the actual achieved figure.
• Species cards are reference information, not a substitute for expert identification in regulated trade.
• Only the six covered species are supported; marine species are out of scope.