RevSearch is a Minimum Viable Product (MVP) showcasing a car reverse image search application. It leverages cutting-edge machine learning, computer vision, and cloud-based technologies to provide an efficient and accurate image search experience. Built as a self-initiative, this project demonstrates end-to-end machine learning workflow expertise.
- Image Upload: Users upload an image in various formats (JPG, PNG, BMP, etc.).
- Feature Encoding: The uploaded image is encoded into a feature vector using a trained neural network encoder (EfficientNet).
- Similarity Search: The encoded feature vector is compared to precomputed feature vectors in the database.
- Top Matches: The system retrieves the top 10 image URLs based on similarity scores.
- Image Retrieval: Top 10 images are fetched from AWS S3 storage via AWS API Gateway and AWS Lambda.
Note: RevSearch is currently not deployed due to associated operational costs.
- Company: Self-initiative project for applying end-to-end machine learning workflows.
- Timeline: April 2022 - May 2022
- Codebase:
- Backend, Frontend, and Core Technology
- Upload images in various formats (JPG, PNG, BMP, etc.) for reverse search.
- Interactive slider to select up to 6 similar images.
- Powered by the EfficientNet neural network architecture for accurate feature extraction.
- Responsive and seamless user experience powered by FastAPI.
- Fully interactive Reverse Image Search WebUI.
| Category | Technologies |
|---|---|
| Core Technologies | Python, PyTorch, ONNX, ONNX Runtime, Pandas |
| Data Preprocessing | Albumentations |
| Model Optimization | MLflow, Optuna |
| Web App & Deployment | Streamlit, FastAPI, Docker, Heroku |
| Cloud Services | AWS S3, AWS API Gateway, AWS Lambda |
| CI/CD & Code Quality | GitHub Actions, Black, Pytest |
| Image Processing | PIL |
- Source: Stanford University AI Lab’s Cars dataset.
- Composition:
- 16,185 images across 196 car classes.
- Serves as the foundation for the feature extractor (encoder).
RevSearch integrates DeepSearchLite, a custom lightweight library, for fast and efficient similarity searches with minimal dependencies.
Find it on PyPi: DeepSearchLite
- Dataset Limitations:
- Cars dataset (16,185 images) is outdated, impacting accuracy for newer models.
- MVP Status:
- Currently demonstrates potential but requires further enhancements.
- Next Steps:
- Expand the dataset to include more images and newer models.
- Refine search algorithms for improved accuracy and speed.
- Incorporate user feedback for additional features and functionality.
This project highlights expertise in advanced technologies and practical solutions for the automotive domain.