An Underwater Target Recognition Method for Sonar Images with High Noise

With the rapid advancement of sonar imaging technology, sonar images have been extensively employed in underwater object detection. However, the inherent high-noise characteristics of sonar imagery often lead to the fusion of target information with ambient noise during feature extraction processes, resulting in the disappearance of critical target features. To address this challenge, this paper proposes a novel Transformer-based model incorporating hybrid-domain attention mechanisms and spatial activation functions for underwater sonar image classification and recognition. The framework first employs a cascade connection of Normalization-based Attention Module (NAM) with Transformer Blocks to extract multi-scale target features from sonar images. Furthermore, we integrate SPD-Conv (Space-to-Depth Convolution) modules to mitigate feature degradation during strided convolution operations. The proposed architecture also implements FReLU (Funnel ReLU) activation functions to enhance spatial awareness during training, effectively addressing the limitation of conventional activation functions that focus solely on local information while neglecting contextual relationships. Experimental results demonstrate that our proposed model achieves a significant improvement in target recognition accuracy for sonar imagery, particularly in long-range detection of small underwater targets. The system exhibits superior performance in preserving discriminative features against complex underwater noise backgrounds, with quantitative evaluations showing a 12.7% increase in mean Average Precision (mAP) compared to conventional CNN-based approaches. This advancement provides an effective technical solution for underwater surveillance systems and marine exploration applications requiring high-precision acoustic imaging interpretation.