Fish Species Recognition using CNN

Fisheries surveys, aquarium monitoring and biodiversity studies all start from the same slow manual step: looking at a photo and naming the fish — work that needs domain expertise and does not scale to large survey collections. Training a convolutional network from scratch on fish photos would demand a large labeled set, so this build uses transfer learning: the model starts from an ImageNet-pretrained backbone (a ResNet-style network) that already recognizes edges, textures and shapes, and only the species classifier head — plus a short full-network fine-tuning phase — is adapted to fish photos. The data comes from the QUT fish dataset / Fish4Knowledge collection (Anantharaman et al.), the standard academic reference for underwater fish recognition. A Flask demo wraps the trained model: upload a fish photo and get the top-3 predicted species with confidence scores. The training notebook logs validation accuracy and loss curves and computes a per-class confusion matrix during the build, so the report's metrics come from the run itself (design target ~85%+).

Dataset & model:
Dataset name: QUT fish dataset / Fish4Knowledge collection (Anantharaman et al.) — the standard academic reference for underwater fish recognition. Source: the public academic release. Task: image classification (fine-grained, single-fish photos). Classes: the freshwater species labels in the dataset (the report documents the exact species list and the stratified train/validation split used in the build).
Model: transfer-learning CNN — an ImageNet pre-trained backbone (ResNet-style, He et al., CVPR 2016) with a fine-tuned species classifier head, trained in two phases (head first, then full-network fine-tuning).
Input: a single fish photo (JPEG/PNG), auto-preprocessed on upload. Prediction: softmax scores over the dataset's species. Output: top-3 predicted species with confidence scores.
Evaluation metrics: validation accuracy, accuracy/loss curves, and a per-class confusion matrix on the stratified validation split — all computed by the training notebook during the build. Design target: ~85%+ validation accuracy — a target for the training run, not a pre-claimed score.

Working:

1. Dataset: fish photos from the QUT / Fish4Knowledge collection are loaded and split into stratified train and validation sets, preserving the species distribution in both.
2. Preprocessing: training images are augmented with flips, crops and brightness/contrast jitter to simulate underwater imaging variation; validation images are center-cropped and normalized.
3. Training phase: an ImageNet pre-trained backbone (ResNet-style) is trained in two phases — the species classifier head alone first, then full-network fine-tuning — with validation accuracy and loss logged each epoch.
4. Evaluation phase: the best checkpoint is selected by validation accuracy; a per-class confusion matrix is computed on the stratified validation split during the build — the report's metrics come from this run, against the design target of ~85%+ validation accuracy.
5. Inference phase (demo): the best .pth checkpoint is loaded by the Flask app at startup; each uploaded photo is preprocessed and passed through the network in a single forward pass.
6. Output: softmax scores are ranked and the top-3 species are displayed with confidence bars; out-of-distribution fish receive the nearest-looking label, which the limitations document.

Specifications:
Model | Transfer-learning CNN — ImageNet pre-trained ResNet-style backbone, fine-tuned classifier head
Dataset | QUT fish dataset / Fish4Knowledge collection (Anantharaman et al.)
Classes | Freshwater fish species per the dataset labels; exact species list documented in the report
Split | Stratified train/validation split, documented in the report
Metric | Validation accuracy on the stratified split, computed during the build; design target ~85%+
Augmentation | Flips, crops, brightness/contrast jitter simulating underwater imaging variation
Input | Single fish photos (JPEG/PNG), auto-preprocessed on upload
Output | Top-3 species predictions with confidence bars
Demo app | Flask web app with photo upload and prediction display
Weights | .pth file shipped with the build

Transfer-Learning Fish Classifier [implemented] — ImageNet pre-trained ResNet-style backbone with a fine-tuned species classifier head, trained on the QUT / Fish4Knowledge collection.
Two-Phase Fine-Tuning [implemented] — Classifier head trained first, then short full-network fine-tuning for domain adaptation to underwater-style imagery.
Underwater-Style Augmentation [implemented] — Brightness/contrast jitter, flips and crops simulate the lighting and turbidity variation of real fish photos.
Training Notebook with Full Logging [implemented] — Validation accuracy and loss curves every epoch, plus a per-class confusion matrix computed during the build.
Demo App with Top-3 Predictions [implemented] — Upload a fish photo and get the top-3 predicted species with confidence scores and bars.
Exportable .pth Weights [implemented] — The checkpoint suits deployment-style demos.
Confidence Threshold and Top-K Display [configurable] — Prediction display settings adjustable in the demo config.
Multi-Fish Detection [future-scope] — Detecting and classifying several fish in one frame is documented as possible future work, not included in this build.

Complete source code: data pipeline, training, evaluation, Flask demo app
Trained CNN weights (.pth)
Dataset download and split-configuration scripts
Project report PDF: background, dataset analysis, methodology, evaluation, error analysis
PPT presentation for final review
Viva Q&A preparation document: transfer learning, CNNs, augmentation, evaluation metrics
Setup guide: environment, dependencies, GPU guidance for training