Claims processing for car insurance can be an administratively costly process as the time investment required from claim adjusters can be intensive. Most expenses incurred by insurance companies are passed on to policyholders, so any process optimizations should be beneficial to both the company and the policyholder.
The main objective of the project was to find the best candidate model to automate the classification of the location of damage on the car in a supplied image. To achieve this, 3 transfer learning techniques are considered were pre-trained models are used as (1) feature extractors, (2) for weight initialization and model fine-tuning then carried out and lastly (3) combination of feature extractor and weight initialization. As part of applying transfer learning to the task, comparison of different optimizers and learning rates is carried out to identify sensitivities of the above techniques.The results identify transfer leaning techniques (2) and (3) to be more sensitive to optimizer and learning rate. The best candidate model for the task selected based on accuracy and F1-score, is a wide residual network.
The full report can be found here and the link to project poster can be found here. The code for implementation can be found in this Github repository.
