In the context of accelerating and automating visual vehicle inspection for insurance and automotive service applications, I developed a proof‑of‑concept deep learning solution to classify images of car fronts and rears into six predefined damage categories. The core challenge was to construct a robust classification pipeline that could generalize across diverse image conditions with limited labeled data and seamlessly integrate into an interactive user interface. The end‑to‑end solution incorporated data preprocessing, model development, deployment, and UI integration to demonstrate both technical proficiency and real‑world applicability.
I implemented Convolutional Neural Networks (CNNs) with transfer learning, primarily leveraging the ResNet50 architecture, to extract and learn hierarchical image features that differentiate between normal and damaged vehicle states. Multiple model configurations were trained and evaluated, including baseline CNNs, regularized variants, and tuned ResNet models. The resulting model with ResNet50 and hyperparameter optimization achieved the strongest performance. For deployment, I built a FastAPI server to expose the model’s predict endpoint and a Streamlit application for interactive demonstrations, enabling users to upload vehicle images and receive real‑time predictions.
The project delivered a functional car damage classifier capable of discerning six distinct damage conditions, backed by a clean codebase on GitHub and a responsive interface for demonstration purposes. The best performing model delivered competitive classification accuracy on the held‑out test set, reflecting sound model selection and tuning practices. Additionally, this work highlights skills in Python software engineering, transfer learning in computer vision, API development, and end‑to‑end ML deployment—each directly reflected in my resume and applicable to roles focused on machine learning engineering and applied AI.