metadata
license: mit
library_name: keras
tags:
- image-classification
- multi-task-learning
- art
- painting-classification
- mobilenet-v2
datasets:
- huggan/wikiart
metrics:
- accuracy
- top-5-accuracy
WikiArt Multi-Task Painting Classifier
A multi-task deep learning model for classifying paintings by artist, genre, and style simultaneously.
Model Description
This model performs three classification tasks on painting images:
- Artist Classification: 129 artists (Claude Monet, Van Gogh, Picasso, Da Vinci, etc.)
- Genre Classification: 11 genres (portrait, landscape, abstract painting, etc.)
- Style Classification: 27 art styles (Impressionism, Cubism, Renaissance, Baroque, etc.)
Model Architecture
- Base Model: MobileNetV2 (pre-trained on ImageNet)
- Framework: TensorFlow/Keras
- Input: 224×224 RGB images
- Approach: Multi-head architecture with shared convolutional base
- Total Parameters: ~3.5M (approximate)
Training Details
Dataset
- Source: WikiArt dataset
- Total Images: 84,440 paintings
- Split: 75% training, 25% validation
Training Procedure
- Preprocessing: MobileNetV2 preprocessing (normalization)
- Augmentation: Random horizontal flip, rotation (±5°), zoom (±10%)
- Optimizer: Adam (1e-3 for frozen, 2e-4 for fine-tuning)
- Loss: Sparse categorical cross-entropy (for all three tasks)
- Training Stages:
- Frozen backbone (2 epochs)
- Full fine-tuning (10 epochs)
Evaluation Metrics
- Top-1 Accuracy (all tasks)
- Top-5 Accuracy (artist and style)