Scene Level Image Classification: A Literature Review

Convolutional neural networks (CNNs) have made significant contributions to natural and remote sensing imaging since the development of deep learning. Scene-level image classification is a challenge that affects both the natural and remote sensing domains and has numerous applications. The number of possible scene entities in the image content that could match the dataset images is the main focus. Scene-level classification is significant and fascinating because of open problems like intraclass heterogeneity, interclass homogeneity, background cluttering, high spatial resolution, and different imaging conditions. Additionally, the multi-label scene dataset’s imbalance, lack of preservation of complex semantic relations, and higher label-to-label correlation are all apparent. The article discusses a meta-analysis of the state-of-the-art scene classification literature practices. We discuss CNNs, attention mechanisms, capsule networks, and generative adversarial networks. The article also delivers an overview of the various activations, losses, optimization techniques, and regularization schemes pertinent to the scene domain. The standard benchmark datasets based on single- and multi-label themes are collated. The performance metrics for scene classification are explained as well. The implementation of the multi-label scene classification utilizing several CNN models on the UC Merced multi-label dataset is also covered in the paper. The proposed MobileNet-based model performs better than the recognized cutting-edge methodologies.