Multiple DeepLofargram: Weakly supervised detection and recovery of multiple dim frequency lines for underwater acoustic targets

Reliable detection and recovery of multiple frequency line structures in Lofargrams are crucial for passive underwater acoustic target analysis. Although deep neural networks have shown promising performance for this task, most existing methods rely heavily on dense pixel-level annotations. To address this limitation, we propose a weakly supervised unified framework for multiple frequency line detection and recovery. Specifically, line detection is conducted under coarse region-level supervision, whereas line recovery is achieved through gradient-based visualization in a weakly supervised manner without pixel-level masks. To ensure that the recovered lines faithfully preserve fine-grained dynamic characteristics, such as continuity and fluctuations, we further introduce a recovery loss that enforces curvature smoothness and structural consistency. Extensive experiments show that the proposed multiple DeepLofargram achieves state-of-the-art performance on multiple coexisting frequency lines with fragmented and fluctuating structures under low-signal-to-noise conditions.