With the advancement of intelligent ship, visible-light visual perception has become essential. However, ships motion often causes motion blur in captured images, which degrades object detection accuracy and poses navigation risks. Existing deblurring methods mainly rely on spatial-domain modeling or multi-scale convolution; however they face challenges such as limited receptive fields, poor frequency-domain utilization, and high computational cost, making them unsuitable for real-time maritime applications. To address these challenges, we propose a lightweight network model (LSFNet) that integrates parallel stripe attention and frequency modulation mechanisms for efficient motion deblurring. Specifically, LSFNet adopts a multi-scale learning strategy to progressively extract and fuse blurry features at different resolutions, enhancing robustness to multi-scale blur. A parallel stripe attention module aggregates neighboring information in both horizontal and vertical directions significantly expanding the receptive field. Additionally, a frequency modulation module decouples features into high- and low-frequency components, emphasizing informative signals via trainable weights, thereby improving detail restoration. Experimental results on the GoPro, RealBlur, HIDE, SMD, and a self-constructed Ocean Visual View Dataset (OVVD) demonstrate that LSFNet achieves superior performance in both deblurring quality and model efficiency, offering suitability practicality and deployment flexibility in intelligent ship systems.