Efficient Outdoor Scene Classification Using MobileNetV3 and Transfer Learning

Abstract

Outdoor scene recognition (OSR) is a basic computer vision task that has been applied in autonomous systems, environmental knowledge, and in mobile vision. Nevertheless, a lot of current deep learning (DL) methods are based on architectures that are computationally-intensive and result in the designs of the systems being unable to be applied to resource-constrained hardware. The given paper is focused on an effective transfer learning (TL) based outdoor scene recognition system based on the MobileNetV3-Small architecture. The given approach consists in the use of the pretrained ImageNet features, where the feature adjustment is used to classify the outdoor environment incurring low computation costs. The experiments were performed on the MIT 8-Scene data, a scenery dataset of eight outdoor categories comprising 2,688 images, which were split on the training, validation and testing subsets. PyTorch (CPU-only) was used to train the model, and it showed that the model is practical in the low-power systems. The proposed method had the highest validation of 92.27% and the test accuracy of 88.51% respectively with a high level of accuracy, precision, and recall, and with the highest level of F1-score among the majority of the scene categories. The robustness and generalization ability of the model is confirmed through qualitative and quantitative measures of the model, such as the confusion matrix analysis and sample prediction. According to the results, the idea of lightweight architectures mixed with transfer learning should be viewed as a reliable method of outdoor scene recognition without having to use a high-end hardware, which makes the set of the suggested types of architecture the one that will be relevant to real-world and embedded uses.