Evaluating Transfer Learning Models for Vehicle Scene Classification Under Adverse Weather Classification Conditions

Abstract

ABSTRACT The unreliability of computer vision systems employed in intelligent transportation systems (ITS) is greatly influenced by the adverse weather. This paper examines how well transfer learning (TL) models can be used to classify weather conditions on vehicle scene images. The proposed framework would use the convolutional neural networks that were initially trained on the ImageNet dataset and subsequently fine‐tuned with the Detection in Adverse Weather Nature (DAWN) data. There are four types of TL architectures, that is, InceptionV3, ResNet50, DenseNet201, and Xception, reviewed to classify images into four possible weather categories, that is, fog, rain, sand, and snow. The experimental outcomes reveal that InceptionV3 model performs the best and is better than the other models by around 2%–3% in performance in terms of classification accuracy. The results indicate that TL has the ability to cope with small amount of data and enhance the stability of visual analysis under poor weather conditions in classification. Such findings can be useful in building effective ITS applications like traffic monitoring system and green transportation systems.