Car Exterior Damage Detection Using Mask R-CNN

Computer Science and Engineering

Recently, the production of image-based vehicle insurance is an important area with considerable scope for automation reach. In this paper we consider the issue of classifying car damage, where some of the categories may be fine-granular. To this reason, we are exploring deep learning-based techniques. Initially we try to train a CNN directly with a collection of training data. However, it's not working well due to a small collection of labeled data. Hence, we investigate the domain-specific pre-training effect accompanied by fine-tuning with a large number of annotated training-data. As Faster R-CNN and SVM have not identified damaged cars with high accuracy, and the Cascade R-CNN takes an immense amount of time to train and check the data that we are working on to fit. Hence, we are training data into a R-CNN Mask that produces adequate results compared to traditional Neural Networks. Though there is a lot of unknowns such as partial images, Hence the classifier was built to detect amorphous damages. The model is layered over 3 classifications of detecting the car and examining whether the damage dealt is high or low. Finally, the classifier is projected with the flask environment to make the working experience easier, since it runs on a localhost the compile time does not exceed 5 seconds irrespective of the quality of the image. Experimental results indicate that Mask R-CNN works better than convolutional R-CNN as transfer learning works better than domain specific fine-tuning such as Cascade and Faster R-CNN. We achieve 89.5 per cent accuracy with Mask R-CNN through transfer combination.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.