Deep Visual Waterline Detection within Inland Marine Environment

Jing Huang, Hengfeng Miao, Lin Li, Yuanqiao Wen, Changshi Xiao

Abstract

Waterline usually plays as an important visual cue for maritime applications. However, the visual complexity of inland waterline presents a significant challenge for the development of highly efficient computer vision algorithms tailored for waterline detection in a complicated inland water environment. This paper attempts to find a solution to guarantee the effectiveness of waterline detection for inland maritime applications with general digital camera sensor. To this end, a general deep-learning-based paradigm applicable in variable inland waters, named DeepWL, is proposed, which concerns the efficiency of waterline detection simultaneously. Specifically, there are two novel deep network models, named WLdetectNet and WLgenerateNet respectively, cooperating in the paradigm that afford a continuous waterline image-map estimation from a single captured video stream. Experimental results demonstrate the effectiveness and superiority of the proposed approach via qualitative and quantitative assessment on the concerned performances. Moreover, due to its own generality, the proposed approach has the potential to be applied to the waterline detection tasks of other water areas such as coastal waters.

Introduction

Detecting the waterline is a fundamental task for maritime navigation, acting as a primary visual cue for obstacle avoidance and localization for Unmanned Surface Vehicles (USVs). While detecting the horizon at sea is relatively straightforward, inland marine environments—such as lakes and rivers—present a significant computer vision challenge. These environments are often plagued by environmental noise (fog, rain, "water glint"), visual complexity (shadows, reflections of buildings, erratic shorelines), and hardware constraints that limit onboard computational power. Traditional methods relying on hand-crafted features like color gradients often fail under changing lighting or weather conditions.

Methodology: The DeepWL Framework

To address these challenges, the authors propose DeepWL, a dual-network paradigm designed for robust and efficient waterline identification.

The DeepWL architecture featuring WLdetectNet for detection and WLgenerateNet for synthetic data generation.

WLdetectNet and the WLpeephole

The core of the system is WLdetectNet, an end-to-end binary classification network. A key innovation is the WLpeephole mechanism. Instead of analyzing isolated pixels, the network utilizes a dual-scale receptive field: an "observing field" ($r \times r$) for context and a smaller "recognizing field" ($s \times s$) where the waterline must reside. This structure forces the model to learn the spatial relationship between water and land textures. To ensure real-time performance on USV hardware, the architecture employs Group Convolutions (PGconv) and Channel Shuffling, achieving a deep 72-layer network with only 3.12 MFLOPs.

WLgenerateNet: Data Augmentation via GAN

To overcome the labor-intensive process of manual labeling, the authors introduced WLgenerateNet, a Generative Adversarial Network (GAN). This "data factory" generated approximately 8,000 synthetic waterline samples to augment a smaller manual dataset of 2,000 samples. This approach significantly improved the model's ability to generalize across diverse environmental conditions.

Experimental Results

The system was evaluated using data collected from a USV on East Lake in Wuhan, China. The final balanced dataset comprised approximately 22,000 samples, including manual, synthetic, and negative patches.

Performance and Robustness

Quantitative analysis showed that DeepWL achieved a peak F1-score of 0.943 using optimal window sizes ($r=60, s=30$). In comparative tests against traditional methods like the Canny edge detector, DeepWL demonstrated superior robustness. For instance, in foggy conditions, while Canny's recall plummeted to 41.9%, DeepWL maintained a recall of 84.5% and a precision of 90.6%.

Qualitative comparison of waterline detection across various conditions: (a) clear, (b) sunset, (c) foggy, and (d) distant views. The top row illustrates the scattered noise typical of traditional edge detection, while the bottom row shows DeepWL's accurate estimations.

Efficiency

The computational efficiency of the model is a highlight, requiring only 3.12 MFLOPs. This makes it highly suitable for deployment on embedded systems typically found on USVs, allowing for real-time video stream processing.

Conclusion

DeepWL provides a robust solution for waterline detection in complex inland environments. By leveraging a dual-network architecture and the novel WLpeephole mechanism, the system maintains high accuracy and efficiency where traditional methods fail. Future work may focus on improving the marking strategy for smoother visual results and incorporating temporal data from video flows to enhance stability in choppy waters.

Summary updated

Abstract

Waterline usually plays as an important visual cue for maritime applications. However, the visual complexity of inland waterline presents a significant challenge for the development of highly efficient computer vision algorithms tailored for waterline detection in a complicated inland water environment. This paper attempts to find a solution to guarantee the effectiveness of waterline detection for inland maritime applications with general digital camera sensor. To this end, a general deep-learning-based paradigm applicable in variable inland waters, named DeepWL, is proposed, which concerns the efficiency of waterline detection simultaneously. Specifically, there are two novel deep network models, named WLdetectNet and WLgenerateNet respectively, cooperating in the paradigm that afford a continuous waterline image-map estimation from a single captured video stream. Experimental results demonstrate the effectiveness and superiority of the proposed approach via qualitative and quantitative assessment on the concerned performances. Moreover, due to its own generality, the proposed approach has the potential to be applied to the waterline detection tasks of other water areas such as coastal waters.
BibTeX
@misc{huang2019deep,
  author = {Jing Huang and Hengfeng Miao and Lin Li and Yuanqiao Wen and Changshi Xiao},
  title = {{Deep Visual Waterline Detection within Inland Marine Environment}},
  year = {2019},
  eprint = {1911.10498},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.10498},
}
Figure