University Indoor Scene Classification using Transfer Learning

Abstract

Image classification is used to classify images using a machine without any human intervention. This will help obtain all possible information of that particular capture just by feeding the image or by just clicking pictures from a device. The proposed study has created an extensive dataset of labs and its amenity by comprising 4502 images. Here there are 4 classes in the dataset: Apple Lab, Sophos Lab, Sophos Rack, and Virtual Reality Lab. The proposed work is based on transfer learning. Data augmentation was performed after the proposed model was passed through VGG 16, ResNet50, Inception V3, and Xception pre-trained models from which features were extracted automatically and categorized into 4 distinct classes. Size of the model, inference time, training accuracy, testing accuracy, recall, precision, F1 score, and FLOPs were evaluated while analyzing through pre-trained models. Best experimental results were obtained using Xception pre-trained model among all 4 pre-trained network models with 99.90% as training accuracy, 99.75% as testing accuracy, 99.35% as recall, 99.85% as precision, and 99.60 % as F1 score; having a model size of 83 MB.