chinesemusk/car-damage-resnet50

---
license: apache-2.0
tags:
  - image-classification
  - tensorflow
  - car-damage
  - resnet50
  - vehicle-inspection
---

# 🛠 Car Damage Detection with ResNet50

This model uses a fine-tuned **ResNet-50** architecture to detect and classify car damage types from images. It predicts six classes:

- dent  
- scratch  
- crack  
- broken_light  
- glass_damage  
- no_damage  

### 🔍 Use Case

Primarily designed for vehicle insurance automation, damage assessment, and fleet monitoring systems. Upload a car image and the model will return the most likely type of damage along with confidence scores.

### 📦 Model Format

Exported in **TensorFlow SavedModel format**, compatible with Keras 3 via `TFSMLayer`.

### 🚀 Inference Example (Google Colab)

```python
# Install dependencies (if not done yet)
!pip install -q tensorflow huggingface_hub

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
from tensorflow.keras.applications.resnet50 import preprocess_input
from tensorflow.keras.preprocessing import image
from huggingface_hub import snapshot_download
from google.colab import files

# Step 1: Download model from Hugging Face
model_dir = snapshot_download("chinesemusk/car-damage-resnet50")

# Step 2: Load the model using TFSMLayer for TensorFlow SavedModel format
model = tf.keras.layers.TFSMLayer(model_dir, call_endpoint="serving_default")

# Step 3: Upload an image from Colab
uploaded = files.upload()
img_path = next(iter(uploaded))

# Step 4: Preprocess image
def prepare_image(img_path):
    img = image.load_img(img_path, target_size=(224, 224))  
    img_array = image.img_to_array(img)  
    img_array = np.expand_dims(img_array, axis=0) 
    img_array = preprocess_input(img_array) 
    return img_array

img_preprocessed = prepare_image(img_path)

# Step 5: Running inference
output = model(img_preprocessed)
logits = output['output_0']  
probs = tf.sigmoid(logits).numpy()[0]  

# Step 6: Interpret results
labels = ['dent', 'scratch', 'crack', 'broken_light', 'glass_damage', 'no_damage']
predicted_class_index = np.argmax(probs)  
confidence = probs[predicted_class_index]
predicted_class = labels[predicted_class_index]

# Print the class probabilities for all labels
print("Class probabilities:")
for label, prob in zip(labels, probs):
    print(f"{label}: {prob:.4f}")

img = Image.open(img_path)  
plt.imshow(img)  
plt.axis('off')  
plt.title(f"Prediction: {predicted_class} ({confidence:.2f} confidence)")  
plt.show()

print(f"\nFinal Prediction: {predicted_class} with {confidence:.2f} confidence")

📸 Input

Images resized to 224x224, RGB format.

📤 Output

A 6-element probability vector (after sigmoid), each corresponding to a car damage class.

---