Intelligent underwater mine detection using multi-backbone feature fusion with temporal attention-gated unit on side-scan sonar imaging

Underwater mines pose a high risk to submarines and ships. As a result, navies across the globe employ mine countermeasure units to counter them. Among the measurements employed by MCM units is mine hunting, which involves examining every mine in a suspicious region. With the development of sonar image technology, these images are broadly used for oceanographic studies in military and civilian applications. High-resolution image sonars may be mounted on several survey platforms, typically autonomous underwater vehicles that give improved speed and enhanced data quality with longer-range support. The classification and detection stages are generally executed employing a sonar mounted on an underwater vehicle or on a ship's hull. After recovering the sonar data, military personnel scan the seabed imaging for detecting the targets and classifying the benign objects or mine-like objects. In recent years, researchers have examined the application of deep learning for the classification and detection of mine-like objects from sonar images. This manuscript focuses on the design of Intelligent Mine Detection and Object Recognition Using Multi-Backbone Feature Fusion (IMDOR-MBFF) approach in Side-Scan Sonar Imaging data. The main objective of the IMDOR-MBFF framework is to effectively improve detection and classification performance for real-time underwater mine detection using autonomous underwater vehicles. Primarily, a gamma distribution-based speckle noise model with logarithmic transformation was deployed as the image pre-processing phase to suppress multiplicative noise and enhance structural details. Subsequently, object detection is performed using a RetinaNet architecture with backbone fusion to accurately localize mine-like targets under complex seabed conditions. In addition, discriminative features are extracted through a lightweight yet powerful fusion of GhostNetV2, ConvNeXt, and ShuffleNetV1 is capturing both fine-grained textures and global contextual information. After feature extractor, a feature fusion strategy based on concatenation has been employed to efficiently integrate the complementary information extracted from the three backbones. Finally, a temporal attention-gated unit was employed for accurate mine classification by modeling temporal dependencies across sonar image sequences. The experimental results of the IMDOR-MBFF framework are evaluated on benchmark datasets, and the comparative analysis demonstrates its superior performance over existing methodologies across multiple evaluation metrics.