Weather classification from images using a hybrid deep learning model

Deep convolutional neural networks are widely used today in the fields of image recognition and classification. In these architectures employed for image recognition, convolutional layers autonomously extract features from images, and classification is carried out using these features. While the convolutional layers of different architectures may highlight certain features, they may also lead to the loss of other information. This situation necessitates ongoing research in the field of image processing. While some studies focus on developing new architectures, others aim to create hybrid structures by combining existing models. Image processing is actively used in many fields today. One of these areas is the development of autonomous vehicles. Weather conditions are a crucial factor for driving safety in autonomous vehicles. Using deep convolutional neural networks, it is possible to detect weather conditions from the images captured by the vehicle. In this study, a classification task was conducted on a dataset consisting of images representing various weather conditions. To ensure the originality of the study, two hybrid models were derived from existing architectural models. In designing the hybrid models, popular architectures such as VGG19, ResNet152V2, MobileNetV2, NASNetMobile, and InceptionV3 were used for feature extraction. The extracted features were classified using both an LSTM network and a fully connected network. The highest classification accuracies were measured as %94.47 for the LSTM network and %94.76 for the fully connected network. These results demonstrate the effectiveness of the proposed hybrid model.