My Guava Leaf Disease Classifier

On-device CNN classification for guava leaf diseases using LiteRT & CameraX

---

Overview

My Guava Scanner is a privacy-first, fully offline Android application that identifies guava (Psidium guajava) leaf diseases using a custom CNN model running on-device via LiteRT (Google's edge AI runtime). Point your camera at a guava leaf — the app classifies the condition in real time with configurable FPS and hardware acceleration.

The classification model is based on the research documented in:

Kumar, R. (2024). Guava Leaf Disease Scanner using Deep Learning and Edge AI.
DOI: 10.13140/RG.2.2.12835.92961

All inference happens entirely on-device — no internet connection required after the initial model download. Your images never leave your phone.

---

Features

• 📸 Real-time classification — Live camera feed with <100ms inference at 20 FPS on modern devices
• 🔬 9 disease classes — Algal Leaf Spot, Black Mold, Red Rust, Scab, Scorch, Insect Bite, Yellow Leaf Disease, plus healthy leaf & fruit
• ⚡ Hardware acceleration — CPU, GPU, or NPU with automatic GPU→CPU fallback
• 📷 Gallery import — Analyze existing photos via the system photo picker
• 🎯 Square crop overlay — Guides optimal framing for the 256×256 model input
• 🔦 Flash & camera flip — Front/back camera toggle with flash control
• 📊 Nerd info panel — Real-time metrics: inference time, memory usage, image dimensions
• ⚙️ Configurable performance — Choose between 10 or 20 FPS classification rate
• 🎨 Material 3 Design — Dynamic color theming with light/dark mode support
• 🔒 Fully offline — Model downloaded once via WorkManager, then zero network access
• 🔄 Exponential retry — Robust download with backoff strategy

Classification Outputs

Label	Display Name	Severity
healthy_leaf	Healthy Leaf	✅ Low
fruit_healthy	Healthy Fruit	✅ Low
algal_leaf_spot	Algal Leaf Spot	⚠️ Medium
insect_bite	Insect Bite	⚠️ Medium
scab	Scab	⚠️ Medium
black_mold	Black Mold	🔴 High
red_rust	Red Rust	🔴 High
scorch	Leaf Scorch	🔴 High
yellow_leaf_disease	Yellow Leaf Disease	🔴 High

---

Architecture

My Guava Scanner follows a multi-module architecture with clear separation of concerns:

---

Navigation Flow

Download ──► Camera ──► Settings
   │                       │
   └── (auto route) ───────┘    (pop back)

---

Tech Stack

Category	Technology	Version
Language	Kotlin	2.3.20
UI	Jetpack Compose + Material 3	2026.03.00 BOM
Navigation	Navigation3 Compose	1.1.0
Camera	CameraX (core, lifecycle, camera2, compose)	1.6.0
Inference	LiteRT (Google AI Edge)	2.1.0
DI	Dagger Hilt	2.57.1
Networking	OkHttp + Retrofit + kotlinx-serialization	5.3.2 / 3.0.0
Async	Kotlin Coroutines	1.10.2
Background	WorkManager + Hilt integration	2.11.2
Persistence	Room + DataStore Preferences	2.8.4 / 1.2.1
Image Loading	Coil Compose	2.7.0
Animation	Lottie Compose	6.6.2
Logging	Timber	5.0.1
Build	Android Gradle Plugin	9.1.0
Minimum SDK	Android 7.0 (API 24)
Target SDK	Android 16 (API 36)

---

Getting Started

Prerequisites

• Android Studio Meerkat or later (with AGP 9.1+)
• JDK 21
• Android SDK 36
• A physical Android device (Android 7+) — the camera and GPU acceleration require real hardware

Setup

1. Clone the repository

   git clone https://github.com/DaniZakaria63/MyGuavaScanner.git
   cd MyGuavaScanner

2. Set the model download URL

   Create or edit gradle.properties and add:

   URI_KUMAR_MODEL_CNN_256="<your-model-url>"

   The model is a TensorFlow Lite file (best_cnn_model256.tflite) hosted at your endpoint or Google Drive. See Model Details.

3. Open in Android Studio

   Open the MyPlant/ directory. Android Studio will sync the Gradle project automatically.

4. Run on device

   Select your connected device and hit Run (Shift+F10). The app downloads the CNN model on first launch via WorkManager.

Build Variants

• Debug — Timber logging enabled, debug tensor output printed to Logcat
• Release — Minification disabled (configurable), proguard rules available

# Build release APK
./gradlew :app:assembleRelease

# Run instrumented tests
./gradlew :ai:connectedAndroidTest
./gradlew :app:connectedAndroidTest

---

Model Details

The classifier uses a custom CNN architecture trained specifically for guava leaf disease classification.

• Architecture: Convolutional Neural Network (custom design)
• Input size: 256 × 256 pixels (RGB)
• Normalization: (pixel - 127.5) / 127.5
• Framework: TensorFlow → LiteRT (TFLite format)
• File name: kumar-v1.0.0-256x256.tflite
• Classes: 9 output classes
• Reference: DOI: 10.13140/RG.2.2.12835.92961

Hardware Acceleration

Accelerator	Code	Fallback
CPU	0	None (default)
GPU	1	Automatic → CPU
NPU	2	None

The GPU fallback mechanism in PlantRepository.getModel() catches CompiledModel.create() failures and retries with Accelerator.CPU, ensuring broad device compatibility.

Inference Pipeline

Camera frame / Gallery image
        │
        ▼
  Bitmap (any size)
        │
        ▼
  Scale to 256×256
        │
        ▼
  Rotate (0/90/180/270°)
        │
        ▼
  Normalize: (RGB - 127.5) / 127.5
        │
        ▼
  LiteRT CompiledModel.run()
        │
        ▼
  Argmax over 9 outputs
        │
        ▼
  GuavaHealthResult (label + severity)

---

Screenshots

---

Testing

AI Module Tests

Instrumented tests in :ai cover:

• Model download from network endpoint (KumarModelDownloaderTest)
• Full inference pipeline with 6 sample images at 4 rotations (PlantClassifierTest)
• Determinism verification (same input → same output across runs)
• Error handling (empty model, invalid labels, corrupted files)
• Label list integrity (9 classes, no duplicates)

App Module Tests

Instrumented tests in :app cover:

• ModelDownloadWorker lifecycle — 16 test cases spanning success, retry, failure at all attempt counts (0–3), boundary conditions, and cache/settings integration
• Fake ModelDownloader with SUCCESS/EMPTY_FILE/CORRUPTED/LARGE_FILE scenarios

Running Tests

# AI module
./gradlew :ai:connectedAndroidTest

# App module
./gradlew :app:connectedAndroidTest

# All modules
./gradlew connectedAndroidTest

---

Configuration

Settings Screen

• Hardware Accelerator: CPU (default) / GPU / NPU — GPU availability depends on GLES_VERSION >= 3.0
• Classification FPS: 10 (balanced) / 20 (performance) — controls how often the camera frame is analyzed

Gradle Properties

Property	Description
URI_KUMAR_MODEL_CNN_256	URL for the .tflite model file

---

Contributing

This project is not open for contributions at this time. See License for usage restrictions.

---

License

MyPlant Personal Use License

Copyright © 2024 WALAWE (Dani Zakaria). All rights reserved.

Personal use is permitted — you may download, install, and use this application for your own personal, non-commercial purposes free of charge.

Any other use requires explicit written permission from the author. This includes, but is not limited to:

• Commercial use, distribution, or monetization in any form
• Modification, adaptation, or creation of derivative works
• Redistribution of the source code or compiled binaries (modified or unmodified)
• Integration into other products or services
• Academic or research publication without attribution

To obtain permission, contact:

Dani Zakaria — dani.zakaria@proton.me

Unauthorized use, distribution, or modification of this software is strictly prohibited and may result in legal action.

---