Current Accuracy (5-shot)
98.2%
Inference Time
223
milliseconds
Dataset Images
0
total images
F1-Score
0.971
State-of-the-art

Few-Shot Performance Surface

Model Comparison (Radar Chart)

Defect Distribution

Training Progress History

Recent Inference Logs

TimePredictionConfidenceStatus

Upload Screen Defect Images

Click or Drag & Drop mobile screen images (JPG, PNG, WEBP)

System will automatically validate if image contains a mobile screen

Labeling Interface

Upload images to start labeling

Dataset Inventory

PreviewFilenameDefect TypeSeverityStatusActions
Loading dataset...

Training Configuration

Training Metrics Dashboard

Current Loss: -
Current Accuracy: -
Best Accuracy: 98.2%
Validation F1: -

3D Feature Space (ViT Embeddings Visualization)

● Scratch
● Crack
● Spot
● Discoloration

Previous Training Sessions

Weekly Progress Reports

Draft

Supervisor Feedback

Data Flow Diagram

Activity Diagram

Sequence Diagram

System Architecture

UI InputViT+PromptsPrototype NetOutput
📸 Image Upload
🧠 Feature Extraction
📊 Prototype Distance
✅ Defect Classification

Project Gantt Chart

Project Documentation

Master's Thesis: Few-Shot Vision Transformer for Screen Defect Detection

Student: Du Nanxing (Ahmadi Abdul Nasir) | ID: 982402190105

Supervisor: Prof. Wei Hao | College: School of Automation and Electrical Engineering

Thesis Completion Checklist

Publication Progress

Account Settings

Theme Preferences

System Backup