A deep learning approach for scoring quality in vehicle grading

The demand for used cars is ever-increasing globally, mainly due to availability and affordability. However, even with such high demand in the used car market, most buyers lack knowledge of the factors that can be used to estimate the quality of a used vehicle before the actual purchase. Thus, buyers often tend to obtain the support of a qualified vehicle inspector to evaluate the quality of the car and find out under which grade the vehicle falls before the transaction. This inspection process is intensive, and the vehicle inspectors frequently make mistakes when allocating a class. The proposed research uses deep learning techniques to mitigate the preconceptions of users during manual inspection of used vehicle sales. In this four-step procedure, at first, an image classification model will be used to determine whether the front, back, and two sides of a vehicle appear in all input photos. If it is satisfied, two parallel operations will be executed. The first parallel operation uses a pre-trained deep learning model to identify the vehicles make and model. The original input images are analyzed in the second parallel task using an object detection model to identify vehicle components. After the second parallel step, an extra deep learning model is used to classify the identified attributes, and damages into eight categories. In the last step, a grade is proposed using the NAMA (National Association of Motor Auctions, UK) vehicle grading method based on the specifics of the damaged component. Results reveal a 94% classification accuracy for vehicle orientation and a 97% classification accuracy for make/model using a bespoke VGG16 model. Further, after numerous model comparisons, component identification, and damage assessment have been obtained with 88.1% and 83.7% mAP (Mean Average Precision) values.