Vision with Insight: Improving CNN Architectures Through Interpretability for Artistic Style Detection

Deep learning models often function as black boxes, limiting understanding of their decision-making processes and opportunities for targeted improvements. This project specifically implement a systematic approach CNN refinement guided by interpretability techniques for classifying artistic styles.

Features

• Art style classification across multiple artistic movements
• Supports detection of AI-generated vs human-created art
• Multiple CNN model architectures:
  • Basic CNN (model_basic.py) -CNN with CBAM (model_attention.py)
  • Enhanced CNN with SPP/CIM (model_filter.py)
  • Ultimate model with both CBAM and SPP/CIM (model_ultimate.py)

Visualization Techniques

• Grad-CAM for visualizing model focus areas
• Convolutional filter visualization
• Feature map visualization
• Filter activation through gradient ascent

Data Structure

The system is designed to work with art images organized by:

• Art style categories (impressionism, renaissance, surrealism, etc.)
• Source type (AI_SD_, AI_LD_, human_) You can download the dataset from the kaggle website: https://www.kaggle.com/datasets/ravidussilva/real-ai-art

Requirements

pandas==1.5.3
matplotlib==3.7.1
scikit-learn==1.2.2
opencv-python==4.7.0.72
tqdm==4.65.0
Pillow==9.5.0
seaborn==0.12.2
torchsummary==1.5.1
torch==2.0.1+cu118
torchvision==0.15.2+cu118
torchaudio==2.0.2+cu118
numpy>=1.22.0

Usage

The project contains multiple scripts that can be run independently:

• Train and evaluate models with the main functions in model scripts
• Generate visualizations using Grad_CAM.py and filter_visualization.py