<p>Sea surface image detection is crucial for accurately identifying maritime targets in complex environments, supporting marine environmental perception and engineering applications. This study proposes a dual-threshold coding feature sampling network (DtCFS-Net) inspired by human visual perception. By constructing an image saliency matrix based on visual attention parameters and integrating saturation, hue, and contour features, the proposed model employs layered encoding, adaptive Gaussian filtering, and coefficient transformation to enhance multiscale saliency detection. A dual-threshold function is introduced to suppress non-maximum high-frequency noise and refine contour extraction. The experimental results show that the proposed DtCFS-Net outperforms existing saliency-based feature extraction algorithms, achieving 63.6% and 18.8% improvements in correlation coefficient and normalized scan-path saliency, respectively, along with superior performance in area under the curve (AUC), shuffled-AUC, peak signal-to-noise ratio, and structural similarity index measure. Integrated into a backbone network, it effectively reduces missed detections and false alarms. Compared with YOLOv10, the proposed model improves mAP₅₀ and mAP_50–95 by 4.71% and 1.23%, respectively. These findings underscore its potential applications in marine engineering, where the integration of multimodal sensor data can enhance navigation, positioning, and path planning for unmanned vessels in the future.</p>