Texture Classification using Vision Transformer (ViT)

This project uses a pre-trained Vision Transformer (ViT) model (google/vit-base-patch16-224) to classify images from the Describable Texture Dataset (DTD) into 47 texture categories.
The model achieves high accuracy through transfer learning and data augmentation techniques.

---

Dataset

Name: Describable Texture Dataset (DTD)
Classes: 47 texture categories (e.g., banded, blotchy, bubbly, etc.)

Structure:

dtd/
└── images/
├── banded/
├── blotchy/
├── bubbly/
└── ...

Setup:

1. Download and extract the dataset from the official source.
2. Place it in your Google Drive:

/content/drive/MyDrive/dtd/images

---

Environment Setup

This code is designed to run on Google Colab with GPU enabled.

Required Libraries:

pip install torch torchvision transformers tqdm

---

Model Architecture

• Base Model: ViT Base (Patch16-224)
• Modification: Final classification head modified for 47 classes instead of 1000.
• Pretrained: Yes, on ImageNet, then fine-tuned on DTD.

---

Training Pipeline

Data Augmentation & Transforms

Training:

• RandomResizedCrop
• RandomHorizontalFlip
• ColorJitter
• RandomRotation
• Normalize (using ViT's mean/std)

Validation & Testing:

• Resize → CenterCrop → Normalize

---

Data Splits

• Train: 70%
• Validation: 15%
• Test: 15%

---

---

Performance

Metric	Value
Best Validation Acc	78.25%
Total Classes	47

---

Evaluation

def evaluate(model, loader):
    # Evaluation logic
    ...

---

Saving & Loading Model

# Save
torch.save(model.state_dict(), "vit_texture_model.pth")

# Load
model.load_state_dict(torch.load("vit_texture_model.pth"))

---

References

• Describable Texture Dataset
• Hugging Face Transformers
• PyTorch