22 August 2018
image classification, learning (artificial intelligence), neural nets, pattern classification, feature extraction, object detection, computer vision, traffic engineering computing, discriminative object parts, key information, authors, traditional convolutional neural network models, confidence score, predicted box, traditional approaches, fine-grained classification, bounding box predictors, network training, original images, cascade methods, cascade model, vision-based classification methods, vehicle-type classification, real-world image, intelligent transportation system, car appearance
In recent years, a number of vision-based classification methods have been proposed. However, a few of them were paid attention to vehicle-type classification in a real-world image, which is an important part of the intelligent transportation system. Owing to the large variances of the car appearance in images, it is critical to capture the discriminative object parts that can provide key information about the car pose. In the authors’ project, the traditional convolutional neural network (CNN) models are modified and experiments are followed as well. The model has two main contributions. First, the output shows a confidence score of how likely this box contains a car for each predicted box, which has some certain advantages compared with other models and is quite different from traditional approaches. Another contribution is the fine-grained classification of the makers and models of a car, which need to train the bounding box predictors as part of the network training. The experiment results show that data enhancement and pre-train of CNNs with original images can classify the vehicle makes and models with a high accuracy of nearly 80%. Cropping images by cascade methods can increase the precision to 86.6%.