Technical Implementation

Detailed architecture and implementation of our AI models

Model Architecture

Core components of our AI system

Base Model

  • ResNet18 - Pretrained on ImageNet
  • Fine-tuned on our specialized dataset
  • Modified final layers for our classification task

Framework

  • Fastai - For simplified training pipeline
  • Advanced callback integration
  • Mixed precision training

Dataset Structure

  • Organized by class folders (incisioni/, fotografie/, etc.)
  • Balanced class distribution
  • Metadata annotations for future expansion
Model Architecture

Training Pipeline

Our end-to-end model development process

Data Loading & Augmentation

  • ImageDataLoaders.from_folder for automatic dataset splitting
  • Data Augmentation:
    • Rotations (±20°)
    • Zoom (up to 1.2x)
    • Brightness variations (±40%)
    • Perspective warping
  • Train/Valid split (80%/20%) with fixed random seed (42)
  • Dynamic image resizing (configurable via slider)

Training Parameters

  • Epochs: Configurable (default: 5)
  • Batch Size: Adjustable (default: 32)
  • Optimizer: Adam with weight decay
  • Learning Rate: 1e-3 with cosine annealing
  • Metrics: Accuracy, Error Rate

HuggingFace Space

We've deployed a demo version of our model on HuggingFace Spaces:

Visit our HuggingFace Space

Our HuggingFace space features:

  • Interactive model demo
  • Example predictions
  • Model card with technical details

Model Evaluation

Performance metrics and validation

Confusion Matrix

Confusion Matrix

Key Metrics

Metric Value
Overall Accuracy 92.3%
Precision (Avg) 91.8%
Recall (Avg) 92.1%
F1 Score (Avg) 91.9%

Class-wise Performance

The model shows particular strength in distinguishing between:

  • Lithography vs. Etching
  • Albumen prints vs. Silver gelatin
  • Woodcut vs. Engraving

Future Improvements

Planned enhancements to our system

Model Architecture

  • Upgrade to ResNet34/50
  • Experiment with Vision Transformers
  • Attention mechanisms

Data Expansion

  • Expand to 50+ print techniques
  • High-resolution scans
  • Multispectral imaging

Deployment

  • API endpoints
  • Batch processing
  • Integration with archival systems

API Documentation & Telegram Bot

Interact with our model programmatically or through Telegram

OpenAPI Documentation

Access our API through the standard OpenAPI specification.

Download OpenAPI Spec

Full technical schema available via Redoc interface.

Telegram Bot

Use our bot to classify images directly through Telegram.

Open MetaSophiaBot

Telegram Bot Preview

Ask questions, upload images, and receive AI insights in real time.