E-VAPORISER — Contraband Image Classifier & Seller Portal

Automated contraband screening for online classified-ads marketplaces. This system flags listing photos showing an e-vaporiser (vape) so prohibited items can be blocked before going live. Features an empirical transfer-learning benchmark comparing ResNet-18 vs ViT-B/16, hard-negative curation against lookalike objects, and a live two-model Gradio seller portal with an automated triage verdict.

A marketplace moderation filter cannot rely on easy image classification. A naive model easily cheats via shortcut learning (predicting background resolution or camera noise rather than contraband). The real job is to tell contraband from benign lookalikes — pens, flashlights, lighters, remotes, and power banks — without introducing a manual moderation bottleneck.

What it does

• 🔍 Contraband detection on marketplace listings — binary visual classifier flagging e-vaporisers in product photos before listings publish.
• 🛡️ Hard-negative robustness — trained on 800 real vape photos balanced against 800 negatives (50% resolution-matched everyday scenes from Imagenette-320 + 50% hard negatives from Open Images: pens, flashlights, remotes, microphones).
• ⚖️ Dual-backbone benchmark — empirical comparison between a classic CNN (ResNet-18, 11.7M parameters) and a Vision Transformer (ViT-B/16, 86M parameters), testing both full fine-tuning and frozen head-only strategies across four identical training pipelines.
• 🎯 Moderation-aware threshold tuning — evaluates precision vs recall tradeoffs rather than raw accuracy alone; threshold tuned on validation data to achieve $\ge 99%$ vape recall.
• 🛍️ Dual-model seller portal (Gradio) — uploads are scored live by both top models side-by-side with confidence bars and a three-tier verdict: Blocked (both agree), Needs Review (models disagree), or Allowed (both clear).

Architecture

listing photo ──▶ [seeded resize + normalisation] ──▶ dual-model inference
                                                               │
                                  ┌────────────────────────────┴────────────────────────────┐
                                  ▼                                                         ▼
                       ResNet-18 (full fine-tune)                               ViT-B/16 (full fine-tune)
                       vape probability: P(ResNet)                              vape probability: P(ViT)
                                  │                                                         │
                                  └────────────────────────────┬────────────────────────────┘
                                                               ▼
                                                  verdict triage engine
                                              ┌─────────────────────────────────┐
                                              │ Both >= threshold  ──▶ BLOCKED  │
                                              │ Disagree           ──▶ REVIEW   │
                                              │ Both < threshold   ──▶ ALLOWED  │
                                              └─────────────────────────────────┘
                                                               │
                           Gradio UI: live confidence bars · agreement indicator · portal review

The machine learning (notebooks/evaporiser_classifier_v2.ipynb)

• Backbones:
  • torchvision.models.resnet18 — lightweight CNN (11.7M params), pretrained on ImageNet-1k.
  • google/vit-base-patch16-224 — Vision Transformer (86M params) via Hugging Face.
• Head Adaptation: Pretrained 1,000-class classifier heads replaced with a fresh 2-class binary head (512 -> 2 on ResNet; 768 -> 2 on ViT). A thin ViTWrapper normalises output logits so both backbones share an identical training and evaluation loop.
• Augmentation (every transform justified): RandomResizedCrop (scale 0.7–1.0) for variable seller framing, RandomHorizontalFlip for arbitrary vape orientation, rotation ($\pm 15^\circ$) for tilted handheld phone shots, and ColourJitter for indoor white-balance variation.
• Honest evaluation (zero test-set leakage): File-level stratified 70/15/15 split seeded across all libraries (SEED 42). Champion models and decision thresholds were selected strictly on validation data; the held-out test set was evaluated exactly once for final reporting.

Held-out Test Results

Configuration	Val Acc	Test Acc	Vape Precision	Vape Recall	Vape F1	Architecture / Strategy
vit_full ★	0.996	0.992	0.992	0.992	0.992	ViT-B/16 · Full fine-tuning (Winner)
resnet_full	0.996	0.988	0.983	0.992	0.988	ResNet-18 · Full fine-tuning (0.996 acc at 0.74 threshold)
vit_head	0.992	0.988	0.992	0.983	0.987	ViT-B/16 · Frozen backbone
resnet_head	0.967	0.971	0.952	0.992	0.971	ResNet-18 · Frozen backbone

Run it

Requires Python 3.10 – 3.12.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python app.py

• Model weights aren't committed to the repository (too large for Git). Produce them by running notebooks/evaporiser_classifier_v2.ipynb in Google Colab (takes ~5 minutes on a T4 GPU), then place the exported best_models/ checkpoints into a local models/ directory.
• Interactive demonstration mode: If launched without .pt checkpoints, app.py automatically launches in interactive demonstration mode so the interface and triage engine can be inspected immediately.
• Sample images: Bundled in demo_samples/pos/ (real vape listings) and demo_samples/neg/ (lookalikes such as pens and flashlights).

Project structure

Path	Role
app.py	Standalone Gradio seller portal application with triage verdict engine
classifier.py	PyTorch inference pipeline, ViTWrapper, model loader, and preprocessing
notebooks/	Main training benchmark (evaporiser_classifier_v2.ipynb), Colab webapp, and executed PDF results
demo_samples/	Sample positive (vape) and negative (lookalike) test images for UI evaluation
requirements.txt	Python dependencies

Tech

Python · PyTorch · Torchvision · Hugging Face Transformers · Gradio · Weights & Biases · Pillow · NumPy

Limitations & further work

• Dataset uses isolated single-product images; real marketplace listings often contain complex cluttered backgrounds or multiple bundled accessories.
• Binary classification flags the entire image; object detection (e.g. YOLO) or segmentation would provide bounding boxes and localize concealed components.
• Multimodal extension combining image tokens with listing metadata (title, price, description) to detect evasion keywords (e.g., "juice pods", "aromatherapy sticks").

License

MIT — see LICENSE.