DeepCar 5.0: Vehicle Make and Model Recognition Under Challenging Conditions

Inspired by multi-agent systems (MASs) and ensemble models, a novel method of vehicle make and model recognition (VMMR) based on a vehicle's front-view images is presented in this paper. By exploiting the attention mechanism in this work, it is demonstrated that most of the features used for classifying a vehicle are extracted from its headlight, grill, scoop, and bumper sections. These areas are designated as the regions of interest (ROIs) in our approach. Contrary to the other methods in which a whole ROI is fed to a deep convolutional neural network, in this scheme, different ROIs are extracted from each image, and then a preprocessing block and a distinctive network are designed for each ROI, which is considered as a single agent. Each agent is then trained separately, and a vehicle's type is determined with the collaboration of these agents and based on the blackboard classification system. Also, a new dataset (DeepCar 5.0) is compiled by using the data from the top 50 automakers. This dataset contains 40,185 images of the front views and the front three-quarters of vehicles in 480 different classes, and all the parts of dataset are labeled manually. The proposed technique is able to achieve the accuracies of 92.14 and 96.72% in the automated and the manual scenarios, respectively; and contrary to the current methods, it can perform flawless classification even when just a portion of a vehicle's front view image is available for processing. The dataset and parts of the code are available at: https://github.com/DeepCar/DeepCar5.0.