Deep_learning_Project

🎨 Fine-Grained Art Style Classification using RegNet + Coordinate Attention + Multi-Head Self-Attention (MHSA)

📌 Overview

This project presents a deep learning framework for fine-grained art style classification using an attention-enhanced RegNetY-032 architecture. The proposed model combines Coordinate Attention (CoordAtt) for directional spatial feature refinement and Multi-Head Self-Attention (MHSA) for global contextual reasoning, enabling the network to effectively capture both local artistic textures and long-range compositional relationships.

A systematic evaluation of multiple attention mechanisms and MHSA insertion positions was conducted to identify the most effective architecture. The final hybrid model achieves an optimal balance between classification performance, computational efficiency, and parameter complexity.

---

✨ Key Features

• Fine-grained classification of 12 art styles
• RegNetY-032 backbone with transfer learning
• Coordinate Attention for directional spatial feature encoding
• Multi-Head Self-Attention (MHSA) for global contextual learning
• Comparative analysis of different MHSA insertion depths
• Evaluation of CBAM and Coordinate Attention integrations
• Performance comparison using Accuracy, Precision, Recall, F1-Score, GFLOPS, and Parameters
• Lightweight hybrid architecture with balanced computational complexity

---

🏗️ Project Pipeline

Input Images
      │
      ▼
Dataset Preparation
      │
      ▼
Data Augmentation
      │
      ▼
RegNetY-032 Backbone
      │
      ▼
Coordinate Attention
      │
      ▼
Multi-Head Self-Attention (After 1024 Layer)
      │
      ▼
Custom Fully Connected Classifier
      │
      ▼
12-Class Art Style Prediction

---

📂 Dataset

• Balanced dataset containing 12 art style categories
• Train / Validation / Test split
• Data augmentation techniques:
  • Random Resized Crop
  • Random Rotation
  • Random Affine
  • Horizontal Flip
  • Color Jitter
  • Random Grayscale
  • ImageNet Normalization

---

🧠 Model Variants Evaluated

Baseline Model

• RegNetY-032

MHSA Variants

• MHSA After 1024
• MHSA After 512
• MHSA After 256
• MHSA After 1024 & 512
• MHSA After 1024 & 256
• MHSA After 512 & 256

Attention Variants

• RegNet + CBAM
• RegNet + Coordinate Attention

Hybrid Architectures

• RegNet + CBAM + MHSA
• RegNet + Coordinate Attention + MHSA

---

🏆 Best Performing Model

RegNet + Coordinate Attention + MHSA (After 1024 Layer)

Metric	Performance
Precision	60.90%
Recall	60.36%
F1-Score	60.46%
Accuracy	60.36%
GFLOPS	3.18
Parameters	21.36M

---

📊 Experimental Results

The proposed hybrid architecture demonstrated the best overall performance among all evaluated configurations. Integrating Coordinate Attention with strategically positioned Multi-Head Self-Attention enabled the model to effectively learn both directional spatial features and global contextual relationships while maintaining moderate computational complexity and parameter efficiency.

---

🛠️ Technologies Used

• Python
• PyTorch
• Torchvision
• NumPy
• Pandas
• Matplotlib
• Scikit-learn

---

📁 Repository Structure

Art-Style-Classification/
│
├── dataset/
├── models/
├── training/
├── evaluation/
├── notebooks/
├── results/
│   ├── confusion_matrix.png
│   ├── loss_accuracy_curve.png
│   └── architecture.png
│
├── train.py
├── test.py
├── inference.py
├── requirements.txt
└── README.md

---

🚀 Future Work

• Improve classification of visually overlapping art styles
• Evaluate on larger and more diverse art datasets
• Explore lightweight Vision Transformer architectures
• Integrate Explainable AI techniques for attention visualization

---

👥 Project Team

• Mythri Kodela
• Murari Vedha Smitha
• Hemanth Kumar Mandala

Faculty Guide

• Rimjhim Padam Singh

---

📄 License

This project is developed for academic and research purposes.