You can deploy this model and start earning money today!

Sathya77/
ViT_MvTec

Name: ViT_MvTec
Availability: InStock

transformers

image-classification

google/vit-base-patch16-224

finetune

apache-2.0

endpoints_compatible

Views

No views yet

🔍 ViT-based Anomaly Detection (MVTec Metal Nut)

This model detects defects in metal nuts using Vision Transformer (ViT) embeddings + kNN anomaly scoring.
It separates good parts from defective ones by analyzing the distribution of anomaly scores. This project uses Vision Transformer (ViT, google/vit-base-patch16-224) to extract robust image embeddings. Fine-tuned it for anomaly detection with a 4-channel input (RGB + edges). Embeddings are scored via k-Nearest Neighbors (kNN), where outliers yield higher anomaly scores. Preprocessing includes 224×224 resizing, Canny edge maps, and normalization. Evaluation uses anomaly score distribution, thresholding (mean+std), and accuracy metrics. Visual results highlight Top-K most normal and most anomalous samples for interpretability.

📂 Dataset

Source: MVTec Anomaly Detection dataset (metal nut subset).
Structure:
- train/ → only good samples
- test/ → mixture of good and defective samples

🏗️ Model Overview

This model is a Vision Transformer (ViT) + k-Nearest Neighbors (kNN) anomaly detection pipeline trained and evaluated on the MVTec Anomaly Detection dataset.Unlike standard supervised classification, this approach learns feature embeddings of normal and defective objects, then applies distance-based anomaly scoring with kNN. The model is designed to detect subtle manufacturing defects such as:

Scratches, Contamination, Surface defects, Structural anomalies
The core idea is that normal samples cluster closely in embedding space, while defective ones appear as outliers with higher anomaly scores.

📊 Example Results

Good Samples (Low Scores ~8–9)

Defective Samples (High Scores ~16–18)

📈 Anomaly Score Distribution

🚀 Usage

python

from transformers import ViTModel
import joblib, numpy as np, torch

# Load model + embeddings
model = ViTModel.from_pretrained("Sathya77/ViT_MvTec")
knn = joblib.load("knn.pkl")
train_embeddings = np.load("train_embeddings.npy")

# Inference
with torch.no_grad():
    outputs = model(pixel_values=batch)
    emb = outputs.last_hidden_state[:, 0, :]
dists, _ = knn.kneighbors(emb)
score = dists.mean(axis=1)