Article Preview
TopIntroduction
Automobile insurance protects the insured party from monetary loss due to accidents and other vehicle-related damages. Insurers pay for the damage to a vehicle in an accident based on claim information (policy information, accident report, photos of vehicle damage) and the assessment provided by the adjuster. Although insurance claim management is a vital component of the business value chain (Porter, 2001), in the United States, significant challenges are encountered in processing automobile damage claims; in 2018, there were 9.13 automobile damage claims per 100 insured vehicles, incurring losses of $63.76 billion (Insurance Information Institute, 2020), while 80,000 American drivers have accidents serious enough to call their insurers (Gardiner, 2020). The effectiveness and efficiency of damage assessments determine the profitability and competitiveness of insurers. The insurers need to be proactive rather than reactive (Wang & Siau, 2019). From a customer service perspective, delays in the claim process can be extremely aggravating to the insured client. However, many automobile damage claims involve assessing images taken during extreme weather, such as hail and hurricanes, which overwhelm the workflow of the insurers’ employees. The cognitive challenges for humans force researchers and practitioners to develop innovative approaches, especially AI-based techniques, to address them (Wang & Siau, 2019). Due to their effectiveness, AI techniques can discover latent knowledge from data, provide more customized problem-solving algorithms, and enhance claim processing efficiency and adaptability.
Recently, new business process optimization techniques adopting less traditional data types, such as image data, have emerged. However, such practices are relatively underutilized by insurers for claim processing. Intense competition has forced insurers to find new ways of enhancing customer satisfaction while reducing costs. Many efforts have been made to improve claim processing speed and quality (Borghesi et al., 1999; Dutra et al., 2003; Gruter, 2008; Qi & Xiao, 2018). In addition to traditional claim management methods, automated claims approaches, such as photo-based estimates using mobile apps, have become increasingly common. In recent years, a few insurance companies, such as Ping An Insurance in China, and Liberty Mutual in the United States, have begun exploring and deploying deep learning–based methods to speed up their claim settlements (Qi & Xiao, 2018; Gardiner, 2020), as these are more capable of handling the large and complex datasets that generally appear in the insurance industry than traditional machine learning models.
In this study, novel deep learning-based approaches for detecting and classifying automobile damage images are developed. This research focuses on identifying scratches, dents to automobile bumpers and bodywork, flat tires, and chips and cracks caused by hail, considering the most common visual damage indicators encountered in insurance claims. Owing to a lack of publicly available images of automobile damage, images are collected using the Google search engine. Augmentation techniques are deployed to increase the sample size for better training results. It is noted that the resulting dataset includes almost all common instances of car damage. This dataset is more representative than 'real datasets' from insurance companies, where multiple cases of a single type of damage can create imbalanced data. Hence, by avoiding imbalanced data, which can cause deep learning and machine learning techniques to perform poorly or even fail completely, the advantages of the deep learning technique developed in this study can be demonstrated better. It can provide a generic high-performance deep learning solution for car damage claims in the insurance industry.