Computer Vision Mobile / Deep Learning 2 Semesters

🌿 Plant Disease Detection App — Comparing CNN Architectures

Point a phone camera at a chili or paddy leaf and get a disease diagnosis. The app is the demo — the research is the head-to-head between CNN architectures, and whether a model trained on lab photos survives real Malaysian field conditions.

1The Pitch — Product + Research in One

🛠️ What you build (the demo)

A mobile-friendly app: snap or upload a leaf photo → disease class + confidence + treatment suggestion card. Runs on-device or via a lightweight Flask API.

🔬 What you research (the grade)

Which architecture gives the best accuracy-vs-size trade-off for mobile deployment, and how much does performance drop from clean lab datasets to self-collected field photos — the classic public-vs-local dataset story.

2Problem Statement & Objectives

Pernyataan masalah: Small-scale farmers identify crop diseases by eye and act late; existing detection models are trained on foreign lab-condition datasets and degrade on local crops photographed in real field lighting.

Build a camera-based leaf disease detection app for a chosen local crop (the MVP).
Benchmark three CNN architectures — MobileNetV3, EfficientNet-B0, and YOLOv8-cls — on the same data splits.
Evaluate cross-dataset generalization: train on public PlantVillage data, test on self-collected field images.

3Benchmark Design — the 3 Experiments

Directly mirrors the strongest FYP pattern: public dataset vs local dataset vs cross-dataset.

Experiment	Setup	What it proves
Eksperimen 1 Public dataset	Train + test all 3 architectures on PlantVillage (80/20 split, same augmentation).	Clean-condition ceiling for each architecture; the accuracy-vs-model-size table.
Eksperimen 2 Local field dataset	Train + test on self-collected field photos (~500–1,000 images, varied lighting/background).	Performance on realistic Malaysian field conditions with a small local dataset.
Eksperimen 3 Cross-dataset	Train on PlantVillage, test on the local field set (no fine-tuning).	The generalization gap — quantifies why local data collection was necessary. Strongest finding for Bab 5.

📦 Dataset plan

Public: PlantVillage subset for the chosen crop (3–5 disease classes + healthy). Local: field photo collection with a documented protocol (device, distance, lighting, time of day) — the collection protocol itself is a methodology contribution. Augmentation: rotation, flip, brightness — reported per experiment.

4Evaluation Metrics

Metric	Why it matters here
Accuracy	Headline number per architecture per experiment — the 3×3 results grid.
Precision / Recall / F1 per class	Confusing two diseases means wrong treatment — per-class recall is the safety story.
Model size & inference time	The mobile-deployment trade-off: accuracy per MB and per millisecond on a real phone.
Confusion matrices	Shows exactly which disease pairs get mixed up and whether field conditions worsen it.

5Report Mapping (Bab 1–5)

Chapter	What goes in it (from this package)
Bab 1 — Pengenalan	Problem statement, objectives, scope (one crop, 3 architectures, 2 datasets), constraints (GPU access, collection season), Gantt chart.
Bab 2 — Sorotan Kesusasteraan	Comparison tables: plant disease detection studies (2019–2026), CNN architecture family comparison, lab-vs-field dataset gap studies.
Bab 3 — Metodologi	Data collection protocol, preprocessing + augmentation pipeline, training configuration, app architecture (camera → model → result card).
Bab 4 — Pembangunan & Pengujian	The 3 experiments, 3×3 metrics grid, confusion matrices, on-device latency table, app screenshots.
Bab 5 — Dapatan Kajian	Recommended architecture with evidence, the generalization-gap finding, limitations (single crop/season), future work (more crops, on-device quantization).

6Tech Stack

PythonTensorFlow / PyTorchUltralytics OpenCVFlask APIFlutter / PWA

🔎 Related works — Perplexity API placeholder

Live literature search (lightweight CNN benchmarks and field-condition plant disease papers, 2020–2026) will be generated here via the Perplexity API once the subscription is active.

⏳ Generate Literature Review (coming soon)

← Back to all FYP ideas · Assignment Helper