Snake Species Identification from Photos using CNN (SnakeCLEF Dataset)

Telling a venomous snake from a harmless lookalike is genuinely hard: an Indian rat snake can pass for a cobra at a glance, and kraits are most active at night when visibility is worst. Field guides help only if the observer already knows what to look at, and generic image classifiers trained on a few clean stock photos collapse on real field photos with cluttered backgrounds, partial bodies and harsh lighting. The SnakeCLEF challenge (CLEF / LifeCLEF, hosted on AIcrowd) turned this into a proper benchmark: its 2021 release contains 414,424 photographs of 772 snake species from 188 countries, sourced from iNaturalist, HerpMapper and Flickr, with a heavy long-tailed class distribution. This project builds the complete fine-grained classification pipeline on that dataset — a curated regionally-relevant species subset, transfer learning from an ImageNet-pretrained EfficientNet-B3 backbone, class-weighted training for the long tail, and a web demo that presents predictions with full probability distributions and species cards. The safety framing is part of the engineering: the demo is explicitly an educational prototype, not a field identification tool.

1. The SnakeCLEF 2021 release is downloaded from AIcrowd; a curated subset of regionally-relevant species is selected, unreadable and non-snake images are cleaned, and a stratified train/validation split is created.
2. Training applies augmentation — random crops, flips, rotations, colour jitter and CutMix — so the model tolerates messy real field photos.
3. The EfficientNet-B3 backbone (ImageNet-pretrained) is fine-tuned with a dropout + dense + softmax head, using AdamW, a cosine learning-rate schedule, label smoothing and class-weighted loss for the long tail.
4. The best weights are checkpointed on validation accuracy; the held-out split is then scored once for top-1/top-5 accuracy, per-class precision/recall/F1 and the confusion matrix.
5. In the web demo, an uploaded or sample photo is EXIF-corrected, resized to 300×300 and normalized identically to training data, then passed through the exported weights.
6. The app renders the predicted species with confidence, the top-5 probability bars, the species information card, and the safety disclaimer on every result.

Tech stack:

• Python 3, PyTorch, torchvision/timm
• scikit-learn, Matplotlib (metrics & plots)
• Jupyter Notebook (training & evaluation)
• HTML5 + CSS + JavaScript (photo-upload web demo)
• SnakeCLEF 2021 dataset (AIcrowd / CLEF LifeCLEF)

Dataset & model details

• Dataset: SnakeCLEF 2021 (CLEF / LifeCLEF, hosted on AIcrowd) — 414,424 photographs of 772 snake species from 188 countries, sourced from iNaturalist, HerpMapper and Flickr (noisy web-scraped subset), plus 28,418 expert-identified test images from private collections and museums. Official split: 347,406 train / 38,601 validation. Heavy long tail: the most frequent species (Thamnophis sirtalis) has ~22,000 images; the rarest (Achalinus formosanus) has ~10. The built-to-order training run uses a curated regionally-relevant subset (design target ~40–60 species) with a stratified split.
• Task: Fine-grained image classification; input = 300×300×3 photo, output = probability distribution over the trained species classes.
• Model: EfficientNet-B3 backbone (~12M parameters) pretrained on ImageNet and fine-tuned, with global average pooling, dropout and a dense softmax head; AdamW optimizer, cosine LR schedule, label smoothing, class-weighted cross-entropy.
• Metrics: Top-1 accuracy ~78–85% and top-5 ~93–96% as design targets for the built-to-order training run, plus per-class precision/recall/F1 and confusion matrix. No accuracy is claimed as measured until the training run is executed for the order.

• [Transfer-learning CNN] EfficientNet-B3 pretrained on ImageNet, fine-tuned on the curated SnakeCLEF subset with a dropout + dense + softmax head — documented layer by layer for the viva.
• [Curated species subset] A regionally-relevant subset (design target ~40–60 species) drawn from SnakeCLEF 2021 with a stratified train/validation split and published class counts, so training is feasible on student-grade GPUs.
• [Per-class probabilities + species card] The demo shows the top-5 probability bars plus a card with venom badge, venom type, region, key identifying features and commonly-confused species.
• [Photo upload demo] Upload any snake photo or pick from 8 built-in sample field photos; the pipeline preprocesses it exactly like training data before inference.
• [Safety-first UI] Every result screen carries the educational-prototype disclaimer — the project treats responsible framing as a feature, not a footnote.
• [Full training notebook] Data curation, augmentation, model definition, AdamW + cosine-schedule training loop, checkpointing and evaluation in one reproducible notebook.
• [Confusion matrix & per-class metrics] Precision, recall and F1 per species, so the viva can discuss exactly which species pairs confuse the model (typically cobra/rat snake, viper/python).
• [Grad-CAM explainability] Activation heatmaps showing which image regions drove each prediction — a strong viva talking point on model interpretability.

• Complete training & evaluation Jupyter notebook (curation, augmentation, training loop, metrics)
• Trained CNN weights with the preprocessing/inference pipeline
• Photo-upload web demo with per-class probabilities and species information cards
• Confusion matrix, per-class metrics and Grad-CAM visualizations
• Project report PDF (background, dataset analysis, architecture rationale, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (CNNs, transfer learning, fine-grained classification, long-tailed data, evaluation metrics)