A proof of concept to use computer vision and deep learning to check whether a car is damaged or not and if damaged check severity and location. Trained a pipeline of convolutional neural networks using transfer learning on DenseNet-201 with Keras and Tensorflow to classify damage. Deployment is done using a web app with Flask, dockers and tensorflow serving. Identified damage location and severity to accuracies of 79% and 71% respectively, comparable to human performance. Trained a pipeline of convolutional neural networks using transfer learning on VGG16 with Keras and Theano to classify damage. Deployed consumer-facing web app with Flask and Bootstrap for real- 5400 time car damage evaluations. Data scraped from Google Images using Selenium, hand-labeled for classification and supplemented with the Stanford Car Image Dataset.
- Blog post - Coming soon!
- Web app - Car Damage Detective - Currently unavailable
- Presentation
Access to the image dataset is made available under the Open Data Commons Attribution License: https://opendatacommons.org/licenses/by/1.0/.
Credit for the Google Images scraper goes to Ian London's fantastic General Image Classifier project.