Intelligent ship driving has emerged as a cornerstone of the global shipping industry’s transition toward automated and unmanned operations. These advanced navigation systems rely on a sophisticated integration of artificial intelligence, high-performance sensors, and big data to manage the complexities of modern maritime navigation. However, the operational environment at sea is characterized by significant challenges that frequently compromise the reliability of visual perception, including variable lighting, severe weather conditions, water vapor scattering, and intense sea surface reflections. These environmental stressors lead to widespread image quality degradation such as color distortion, diminished contrast, and blurred target details. Furthermore, the inherent variability in target sizes—from small navigation aids to large vessels—combined with frequent occlusion, complicates the precision of real-time target recognition. To address these critical issues, a recent research study published in the journal Human-Centric Intelligent Systems introduced a target detection framework that synergistically combines wavelet-based image enhancement techniques with an optimized version of the You Only Look Once version 8n architecture.
The proposed framework adopts a dual-branch wavelet-based image enhancement (WIBE) approach designed to improve both global contrast and local texture expressiveness before passing images to the detection model. In the initial processing phase, the system performs adaptive color restoration. Recognizing that blue light attenuates more rapidly due to its shorter wavelength and scattering by atmospheric aerosols, the researchers avoided simple global compensation, which often leads to chromatic aberration. Instead, the model selects the channel with the largest average pixel value as the reference for data compensation. By stretching the compensated channel pixel values and mapping them to expand the distribution range, the system stabilizes the baseline color features. Additionally, the model utilizes a color attenuation map as a weight parameter for local detail compensation, significantly reducing color difference while mitigating the risk of information loss in less attenuated regions.
The image enhancement process continues with a dedicated contrast enhancement module that transforms the RGB image into the CIE 1976 LAB color space. The system specifically targets the L (lightness) channel, dividing its grayscale histogram into discrete intervals based on an entropy-based calculation that optimizes grayscale thresholds. By performing linear mapping across these intervals, the framework achieves a uniform grayscale distribution, which is essential for handling images affected by non-uniform lighting conditions. Meanwhile, for detail enhancement, the researchers employed a multi-scale Gaussian difference approach. By using three Gaussian kernels with standard deviations of 5, 60, and 120, the model captures blurred images containing small, medium, and large-scale information, respectively. These difference maps are synthesized into a comprehensive enhancement template through weighted fusion, effectively suppressing large-scale noise while highlighting fine textures such as ship outlines and wave edges that are often lost in poor maritime conditions.
Once the local detail and global contrast enhancements are generated, the WIBE framework utilizes a wavelet transform to fuse these disparate components. The images are decomposed into low-frequency components (LFC), which contain structural and brightness information, and high-frequency components (HFC), which house horizontal, vertical, and diagonal texture data. The fusion strategy treats LFCs and HFCs independently, applying a weighted average for the structural information and a selection mechanism based on absolute value comparison for the edge-rich detail components. Finally, a wavelet inverse transform (WIT) reconstructs the enhanced image, providing a visually superior input for the downstream target detection network. This frequency-domain collaborative approach is specifically engineered to preserve global stability while refining the local features necessary for accurate ship classification.
Building upon the enhanced image quality, the research team developed a feature-enhanced version of the YOLOv8n detector, integrating the Convolutional Block Attention Module (CBAM) to refine target identification. The architecture of this improved YOLOv8n model is specifically tailored for maritime hardware, which faces strict power consumption and space constraints. The backbone network uses a series of Conv-BatchNorm-SiLU (CBS) and C2f-Faster stacks for multi-scale feature extraction. To increase computational efficiency, the standard Bottleneck structures were replaced by Faster Block units. These units follow a partial convolution workflow that expands and compresses feature channels to balance expressive capacity with processing speed. By replacing computationally intensive convolutions, the model significantly reduces redundant calculations while maintaining the robustness of its feature representations.
The neck network of the detection model incorporates three DySample units, which provide an efficient, adaptive sampling mechanism that eliminates the need for high-resolution guide maps or specialized software packages. This allows the framework to function effectively within resource-constrained onboard systems. Furthermore, the model includes a custom detection head specifically optimized for 4×4 pixel sizes, enabling it to detect small, distant objects that were often missed by standard YOLOv8 models. This small-object detector head concatenates feature information from early and late stages of the detection pipeline, ensuring that fine-grained information is preserved for localization. The integration of CBAM throughout the neck network introduces spatial and channel-wise weight modifications, which efficiently direct the network to ignore background clutter, such as sea waves or water vapor, and focus on high-value features.
To guide the learning process effectively, the model uses a hybrid loss function that combines normalized Wasserstein distance (NWD) and Inner-IoU. The NWD component is particularly effective for small objects, as it treats bounding boxes as Gaussian distributions to mitigate the sensitivity of standard intersection-over-union metrics to positional shifts. In contrast, the Inner-IoU component uses small bounding boxes for high-IoU samples to fine-tune localization and larger bounding boxes for low-IoU samples to match the target’s true scale. This hybrid approach ensures that the model learns effective localization in complex, noisy marine environments where traditional loss functions might struggle to converge on accurate bounding box coordinates.
The experimental evaluation was conducted using the SeaShips dataset, which contains 7,000 high-resolution images and over 31,000 ship target instances under varying sea states and weather conditions. The WIBE image enhancement model achieved a maximum peak signal-to-noise ratio (PSNR) of 19.8, which was 3.2 higher than the second-best generative adversarial network-based comparison model, FUnIE-GAN. In terms of structural fidelity, the WIBE model achieved a structural similarity index measure (SSIM) of 0.74, converging significantly faster than existing benchmarks. When tested under specific weather conditions like fog, the framework’s Color Image Quality Evaluation (CIQE) metric reached 0.672, demonstrating its generalization capabilities and superiority over alternative methods like Ucolor and WaterNet in restoring image clarity.
An ablation study conducted on the WIBE component revealed that the wavelet inverse transform was the most critical element for success; removing it resulted in a 2.4 decrease in PSNR and a 0.11 drop in SSIM, highlighting the importance of the final frequency-domain reconstruction. Similarly, the removal of the detail enhancement module caused a notable 0.13 reduction in image quality metrics, confirming that multi-scale sharpening is essential for preserving fine texture details. These results suggest that the integration of frequency sub-band separation and weighted fusion is a highly effective methodology for improving visual perception in environments plagued by dynamic lighting and low-visibility conditions.
In the target detection experiments, the improved YOLOv8n model demonstrated a significant boost in performance compared to baseline models. The detector achieved an accuracy of 85.9%, which was 7.5% higher than the standard YOLOv8n and substantially outperformed other models like PP-YOLOE-s and NanoDet-Plus. Perhaps more importantly for real-time applications, the improved model maintained a detection speed of 32 milliseconds, which was 22 milliseconds faster than the second-best model in the study. These figures indicate that the lightweight architectural changes and the optimized loss function successfully provided a high-precision, low-latency solution suitable for the real-time demands of intelligent ship driving.
The ablation studies for the detection network confirmed that every proposed component provided a positive contribution. Simply introducing the CBAM attention mechanism increased the model’s mAP@0.5 to 0.85 by suppressing background noise. Further gains were realized by implementing C2f-Faster, which reduced execution time by 15 milliseconds, and the addition of DySample, which boosted accuracy by 1.4%. The final synergy of these components, particularly the combined NWD and Inner-IoU loss function, achieved the highest detection recall of 87.2% and a mean average precision (mAP@0.5) of 0.91. This holistic optimization, from pre-processing to detection, provides a unified framework that improves the overall safety and reliability of intelligent maritime navigation.
Despite these improvements, the authors noted that the model still faces challenges when navigating in extreme, adverse weather conditions. The performance degradation in such environments suggests that while visual perception is significantly improved, there remain limitations to solely relying on optical data. The research team proposed that the resilience of target detection in the future could be further enhanced by fusing multi-modal data, such as incorporating infrared sensors and radar signals alongside optical images. Such a multi-modal approach would likely complement the current visual framework by providing reliable detection when visibility is reduced beyond the capabilities of current image enhancement techniques. Nonetheless, the current framework serves as a robust foundation for future research, particularly as a pre-processing module for higher-level foundation models or as a stand-alone, lightweight solution for onboard embedded systems.
Subject of Research: Maritime Visual Perception and Target Detection for Intelligent Ship Driving
Article Title: A Robust Maritime Visual Perception Framework via Dual-Branch Wavelet Fusion and Feature-Enhanced YOLOv8n
Article References: Luo, G. (2026). A Robust Maritime Visual Perception Framework via Dual-Branch Wavelet Fusion and Feature-Enhanced YOLOv8n. Human-Centric Intelligent Systems, 6(3), 462-473. https://doi.org/10.1007/s44230-026-00167-0
Image Credits: AI Generated
DOI: 10.1007/s44230-026-00167-0
Keywords: Intelligent ship driving, Maritime target detection, YOLOv8n, Wavelet fusion, Image enhancement, Computer vision, Marine navigation safety, Convolutional block attention module, Real-time object detection, SeaShips dataset, Image quality restoration, Small object detection, Visual perception systems, Autonomous navigation, Sensor fusion, Feature enhancement, DySample, Normalized Wasserstein distance, Maritime visual perception framework
Cite Scienmag News
Elena Sutton. (November 17, 2026). Maritime Visual Perception Framework Boosts Detection Accuracy in Complex Marine Environments. Scienmag. https://scienmag.com/maritime-visual-perception-framework-boosts-detection-accuracy-in-complex-marine-environments/
Elena Sutton. "Maritime Visual Perception Framework Boosts Detection Accuracy in Complex Marine Environments." Scienmag, 17 November 2026, https://scienmag.com/maritime-visual-perception-framework-boosts-detection-accuracy-in-complex-marine-environments/. Accessed 17 November 2026.
Elena Sutton. "Maritime Visual Perception Framework Boosts Detection Accuracy in Complex Marine Environments." Scienmag. November 17, 2026. https://scienmag.com/maritime-visual-perception-framework-boosts-detection-accuracy-in-complex-marine-environments/








