src/app/README.md

Vehicle Damage Classification Using Computer Vision

This project is developed as part of the MIS453 Midterm Project.
The goal is to build an end-to-end machine learning application that classifies vehicle damage images into predefined categories and serves predictions via a backend API.

---

📌 Problem Definition

Given a single RGB image of a vehicle, the system predicts one damage class among the following:

• F_Normal
• F_Breakage
• F_Crushed
• R_Normal
• R_Breakage
• R_Crushed

The task is strictly multi-class image classification.

---

📊 Dataset

• Source: DrBimmer / Comprehensive Car Damage (Hugging Face)
• Total samples: 2300 images
• Split: Stratified Train / Validation (80% / 20%)
• Classes: 6 (verified programmatically)

The dataset was visually inspected and class distributions were analyzed before model training.

---

🧠 Model & Training

• Architecture: ResNet18
• Training mode: Offline (no pretrained weight download)
• Loss: CrossEntropyLoss with class weighting
• Input size: 224 × 224
• Device: CPU
• Output: Trained model artifact saved to disk

Artifacts generated after training:

• artifacts/model.pt
• artifacts/label_names.json
• artifacts/confusion_matrix.pt

---

📈 Evaluation

• Validation Accuracy: ~0.57
• Metrics: Confusion matrix + class-wise Precision / Recall / F1
• Strong performance on Front classes, lower recall on Rear damage types (expected due to visual similarity and class imbalance).

The goal of this project is not accuracy maximization, but demonstrating a correct and reproducible ML pipeline.

---

🚀 Backend API (FastAPI)

A FastAPI backend is implemented to serve predictions using the trained model artifact.

Available Endpoints

• GET /health
  Returns API status and class information.

• POST /predict
  Accepts an image file and returns:

  • predicted class
  • confidence score
  • top-3 predictions

---

▶️ Local Setup & Run

1. Install dependencies

pip install -r requirements.txt