A lightweight dual-branch semantic segmentation network for enhanced obstacle detection in ship navigation

Semantic segmentation is essential for ship navigation as it enables the identification and understanding of semantic regions, thereby enhancing the navigational capabilities of smart ships. However, current deep learning techniques encounter challenges in balancing model size and segmentation accuracy due to the complexity of water surface features. In response, we propose a novel lightweight dual-branch semantic segmentation network. The model initially utilizes a specially designed dual-branch backbone to independently extract local details and global semantics from water surface images. The detail branch compresses and reconstructs feature information to mitigate interference from water dynamics, while the semantic branch efficiently expands the receptive field to capture global object relationships. Additionally, we introduce an aggregation module that holistically guides the feature responses to facilitate the sufficient aggregation of dual-branch information. Furthermore, a cascaded fusion approach is proposed to restore diminished localization precision, while also ensuring fusion accuracy by leveraging the segmentation attributes of deep features. Experimental results on visible light datasets from real navigation scenarios demonstrate that our network achieves approximately a 10% improvement in obstacle detection precision compared to existing advanced maritime models. Moreover, within the domain of the latest lightweight and real-time research, our network attains an optimal balance among accuracy, parameter efficiency, and real-time performance. This contributes to enhancing the navigation safety of intelligent vessels and promotes adaptability for onboard deployment.