Vision Mobile is an Android application that performs object detection on images using a TensorFlow Lite model.

Users can select images from their device (Gallery or Camera), view detected objects and their confidence score, manage prediction history, and customize the applications appearance and model settings.

Application screens:

1. Home
   • Starting point of the application
   • Provides navigation to all other screens
2. About
   • Information about the app, features, version and license
   • Describes how the application works and current limitations
3. Settings
   • Configure the application theme (light/dark)
   • Adjust model-related parameters
4. Recent Predictions
   • Displays the latest 10 predictions
   • Each entry includes the image, detected objects and date time of prediction
   • Can be cleared from the Settings screen
5. Results
   • Shows the detection results on a selected image
   • Accessed after choosing an image from the Gallery or Camera

This project is part of my thesis titled: Design and development of an object recognition system using Artificial Intelligence. Related project: Vision Backend

The model currently supports 80 objects.

A full list can be found here.

Vision Mobile is built with modern Android development practices:

• MVVM (Model - View - ViewModel) architecture for clear separation of concerns
• Jetpack Compose UI using Material Design 3 principles
• Reactive, state-driven UI with clean, maintainable code

• Android Studio - Development Environment
• Kotlin - Primary Programming Language
• Jetpack Compose - Moden declarative UI framework

• DataStore - Persistent user preferences
• Hilt - Dependency Injection
• Coil - Image loading and caching
• Room - Local SQLite database
• TensorFlow Lite - On-device deep learning inference

• Google Material Icons
• Google Material theme builder
• Illustrations by unDraw

A major limitation is the reduced accuracy when detecting small, distant, or low-resolution objects. More details discussed in the related Vision Backend project.

Improvements that can be made:

• Real-time object detection (Video inference)
• Automated model update pipeline
• Human keypoint (pose) detection
• Support for multiple object detection models (e.g., high accuracy, lightweight, or domain-specific variants)

• Install the APK provided in the apk directory
• Run from Android Studio