Bird Species Classification using EfficientNet

Telling 200 similar-looking bird species apart takes years of field practice: a sparrow is not a wren, and the differences hide in beak shape, wing bars and posture. Beginners, students and citizen scientists lack an accessible tool that identifies a bird from a photograph and shows how confident it is. Generic image search is inconsistent, and manual identification from field guides is slow and error-prone for non-experts. This is exactly the fine-grained classification problem — one of the hardest in computer vision — where classes differ only in subtle visual cues and the training strategy matters as much as the model. An automated classifier trained on a large annotated set closes the gap: upload a photo, get ranked species predictions with confidence, and learn what distinguishes them. For students, the project is valuable precisely because it is difficult — transfer learning, augmentation and gradual unfreezing become genuinely necessary, and the per-class error analysis gives examiners plenty to discuss.

Frequently asked questions

1. Which dataset is used? CUB-200-2011 (Caltech-UCSD Birds-200-2011): 11,788 annotated photographs across 200 species, the standard academic benchmark for fine-grained classification. The report documents the split and class distribution used in the build.
2. Which model is used? EfficientNet-B0, fine-tuned from ImageNet weights with the classifier head replaced by a 200-way output layer and deeper blocks gradually unfrozen at a lower learning rate.
3. Is the accuracy guaranteed? No. The design target is roughly 75-80% top-1 accuracy; the training notebook computes the actual top-1 and top-5 accuracy on a held-out split, and the report documents the procedure and error analysis honestly.
4. Why EfficientNet-B0 instead of a bigger model? Compound scaling gives the best accuracy-per-parameter of its generation, so B0 trains fast, fits on modest GPUs and infers in well under a second on CPU — practical for a student build and demo.
5. Is this project suitable for a final-year project? Yes. It suits B.E./B.Tech students in Computer Science, AI/ML and Data Science, demonstrating transfer learning, fine-grained classification, augmentation, top-k accuracy evaluation and a Flask deployment.
6. What will I receive? Complete source code, trained .pth weights, dataset preparation and stratified-split scripts, project report PDF, PPT presentation, viva Q&A document and a setup guide.

Dataset & model:
Dataset name: CUB-200-2011 (Caltech-UCSD Birds-200-2011).
Source: publicly released academic benchmark for fine-grained classification.
Task: 200-class fine-grained image classification.
Classes: 200 North American bird species (11,788 images total), each with common and scientific names.
Model: EfficientNet-B0, fine-tuned from ImageNet transfer learning with gradual unfreezing.
Input: bird photograph, resized to 224x224 and normalized.
Prediction: 200-way softmax over species classes.
Output: top-5 ranked species with confidence bars, plus the predicted bird's fact card.
Evaluation metrics: top-1 and top-5 accuracy, loss curves, confusion matrix, per-class accuracy report — computed by the training notebook on the held-out split during the build.
Design target: ~75-80% top-1 accuracy on the held-out split. This is a design target, not a measured claim; the notebook computes the actual scores during the build.

Working:

1. Training phase: CUB-200-2011 images are split into train, validation and held-out splits with stratified sampling; training images are augmented each epoch with random crops, horizontal flips and color jitter.
2. An EfficientNet-B0 pre-trained on ImageNet is loaded and its classifier head is replaced with a 200-way output layer; the new head is trained first, then deeper blocks are gradually unfrozen with a lower learning rate.
3. Evaluation phase: the training notebook computes top-1 and top-5 accuracy on the held-out split, logs loss curves and builds a confusion matrix with a per-class accuracy report.
4. Inference phase: the Flask app resizes each uploaded photo to 224x224, normalizes it and runs one forward pass through the trained model.
5. The top-5 species with confidence bars are displayed alongside a fact card (common name, scientific name) for the predicted bird.
6. Photos similar to the training data (daylight, reasonably full bird in frame) classify best; blurry or partial birds are reported honestly as less reliable in the demo guidance.

Specifications:
Model | EfficientNet-B0, fine-tuned (ImageNet transfer learning)
Dataset | CUB-200-2011: 11,788 images, 200 species
Metric | Top-1/top-5 accuracy — logged by the training notebook on the held-out split during the build
Design target | ~75-80% top-1 accuracy (design target, not a measured claim)
Input | User-uploaded bird photo, auto-resized to 224x224
Inference | Well under a second per image on CPU (design target)
Demo app | Flask web app with upload UI and top-5 results
Weights | .pth file shipped with the build

[200-Class Bird Species Classification] (implemented) — The fine-tuned EfficientNet-B0 classifier predicts the species of any uploaded bird photo across all 200 CUB-200-2011 classes.
[Top-1 and Top-5 Predictions] (implemented) — The demo shows the single best guess plus the top-5 ranked species with confidence scores, which is more informative when near-identical species compete.
[Prediction Visualization] (implemented) — The top-5 species are ranked with confidence bars and the predicted bird's fact card (common name, scientific name) from the class list.
[Species Fact Cards] (implemented) — Each predicted species displays its common and scientific name drawn from the dataset's class list, making the demo informative for non-experts.
[Data Augmentation Pipeline] (implemented) — Random crops, horizontal flips and color jitter are applied to training images each epoch to reduce overfitting on the fine-grained classes.
[Training & Evaluation Notebook] (implemented) — Logs top-1/top-5 accuracy and loss curves during training and computes a confusion matrix and per-class accuracy report on the held-out split.
[CPU-Speed Inference] (implemented) — The lightweight B0 model infers in well under a second per image on a regular laptop CPU, so the review demo needs no GPU.
[Configurable Top-k Display] (configurable) — The number of displayed ranked predictions and confidence formatting can be adjusted in the app settings.

Complete source code (training, evaluation, inference, Flask app)
Trained EfficientNet-B0 weights (.pth)
Dataset preparation and stratified-split scripts
Project report PDF (dataset background, methodology, evaluation, error analysis)
PPT presentation for the final review
Viva Q&A preparation document (CNNs, transfer learning, fine-grained classification, top-k accuracy, compound scaling)
Setup guide (environment, dependencies, dataset download steps)

CUB-200-2011 covers North American species photographed mostly in daylight — birds outside this set, or in poor light, are misclassified more often.
Fine-grained classes (near-identical sparrows, gulls) remain the main error source; the report documents per-class confusion honestly.
Cluttered backgrounds and partial birds degrade results.
All reported figures are design targets from the buyer's own training run; no pre-measured accuracy is claimed.
This is an educational prototype, not a substitute for expert ornithological identification.