DFF-Net: An Efficient Target Detection Network With Deep Feature Extraction for Underwater Sonar Images

Abstract

Vision-based perception models have been widely used for target detection in underwater sonar imagery. However, due to environmental constraints and hardware limitations, existing models often struggle to capture discriminative deep features in low-resolution sonar images. To address this issue, we propose a novel deep learning (DL) model termed deep feature focusing network (DFF-Net). Tailored to multiscale sonar targets, it strengthens feature representation and captures boundary details alongside high-level semantics. First, DFF-Net incorporates an atrous spatial pyramid aggregation (ASPA) module into the backbone, which jointly models multiscale receptive fields and sonar structural information. It reduces target detail degradation and improves discriminative ability in complex scenarios. Building upon this, DFF-Net further develops a convolution-enhanced multihead self-attention (CEMHSA) mechanism within the backbone. This module combines global dependencies and local information for unified representation. It enhances deep feature representation to capture intrinsic sonar target features precisely. Finally, DFF-Net adopts a hybrid structure-aware contrastive learning mechanism. Utilizing a dynamic weight coefficient allocation strategy, the network achieves complementation between middle-layer features and classification features. It reduces false negatives while steadily improving the generalization ability and environmental adaptability of the model. Extensive experiments on two distinct sonar image datasets verify the competitive overall performance of the model under mainstream evaluation metrics.