SI-GAT: Enhancing Side-Scan Sonar Image Classification Based on Graph Structure

Underwater target classification based on side-scan sonar (SSS) imagery holds significant military and civilian applications. However, challenges, such as low sonar image resolution and limited target features, pose substantial difficulties in the classification of SSS images. The existing deep-learning-based methods for SSS image classification primarily focus on the bright areas of targets, often neglecting information from acoustic shadow regions, global features, and spatial context features. To address these challenges and bolster the correlation among regions containing valuable information, thus enhancing the classification accuracy of SSS images, we propose the SI-GAT model, a graph-based classification network tailored for SSS images. Initially, simultaneous feature extraction is performed on both target bright areas and shadow regions to obtain more comprehensive target features. Subsequently, a weighting function and a metric function are defined to calculate feature distances and correlation matrices, capturing the spatial relationships of targets. The introduction of the K-nearest neighbors (KNN) algorithm and attention mechanism adaptively allocates aggregation coefficients within the neighborhood, promoting the aggregation and propagation of features. Finally, the features of all nodes in the aggregated graph are combined to extract global features. Through testing on a real SSS dataset and comparing the convergence and classification performance with state-of-the-art CNN and GNN models, the results validate the effectiveness of our proposed method.