Bag-of-Visual-Words Feature Extraction using Graph Convolutional Networks

Feature representations are the fundamental building blocks for various image analysis tasks like retrieval and classification. Traditionally, the Bag-of-Visual-Words (BoVW) framework has been a well-established approach for feature extraction. However, BoVW exhibits limitations in capturing the complex spectral-spatial relationships and fine-grained details within remote sensing images (RSIs). This paper proposes a novel deep learning approach that addresses these shortcomings and fits well with the BoVW framework for RSI analysis. Our method leverages a hybrid architecture that combines a Convolutional Neural Network (CNN) and a Graph Convolutional Network (GCN). The CNN extracts high-level spatial features from RSIs capturing prominent visual patterns and spatial arrangements within the image. These extracted features are then fed into the GCN, which operates on a graph structure constructed from the image data. The GCN excels at learning complex relationships between image elements based on their spatial proximity and spectral information. Processing the CNN outputs through the GCN aims to extract and enhance the features that encapsulate both the spatial layout and the intricate spectral-spatial relationships within RSIs. This potentially translates to improved performance in various RSI analysis tasks that demand a nuanced understanding of the image content.