RT-FogNet: real-time ship detection under low-visibility conditions in inland waterways

Abstract

Real-time ship detection in inland waterways is critically challenged by low-visibility conditions such as dense fog, backlight, and surface reflections. To address this, we propose RT-FogNet, a lightweight and deployable detection framework tailored for practical inland waterway scenarios. RT-FogNet integrates a Water Surface Image Dehazing (WSID) module to improve image clarity through a domain-adaptive compression and alignment strategy, and builds upon an enhanced YOLOv10-based detection backbone, into which a Dilated Shared Spatial Pyramid Pooling Fast (DS-SPPF) module is incorporated to strengthen multi-scale feature representation with minimal computational overhead. To support robust evaluation under diverse weather conditions, we introduce the Inland Waterway Ship Dataset (IWSD), comprising 9,305 annotated images collected from real-world inland scenes. Experiments conducted on COCO, SeaShips, and IWSD datasets demonstrate that RT-FogNet achieves significant performance gains, including over 15 percentage points improvement in AP on IWSD, while maintaining high inference speed and low parameter complexity. These results confirm the practical suitability of RT-FogNet for real-time deployment in vision-based ship monitoring systems. The dataset and code are publicly available at https://github.com/Object-Detection-01/RT-FogNet.

Introduction

Real-time ship detection is a fundamental component of intelligent inland waterway monitoring systems, as it supports navigation safety, traffic scheduling, and vessel management. Compared with open-sea environments, inland waterways such as rivers, lakes, and artificial channels present more complex visual conditions. Narrow navigation spaces, cluttered backgrounds, and frequent visibility degradation pose persistent challenges to reliable perception. Adverse factors, including fog, strong glare, and low illumination, often degrade image quality and significantly complicate robust ship detection in practical deployments¹.

Among different sensing modalities, vision-based systems such as fixed surveillance cameras provide a cost-effective and easily deployable solution for inland monitoring. They also serve as a practical complement to radar and Automatic Identification Systems in infrastructure-constrained regions². While deep learning has achieved remarkable progress in general object detection, its application to ship detection under low-visibility inland conditions remains limited³. This gap highlights the need for perception systems that are both lightweight and robust to visual interference, particularly in foggy and low-contrast scenarios.

Recent studies have focused on lightweight detectors tailored for embedded or edge devices. Yue et al. have proposed STD-YOLOv5 to reduce model complexity while maintaining detection accuracy⁴. Sun et al. have incorporated ShuffleNetV2 into YOLOv5 and have achieved more than 70% parameter reduction⁵. Gao et al. have developed a compact variant of YOLOv8 for ship detection under normal visibility conditions⁶. Despite their efficiency advantages, most existing detectors exhibit notable performance degradation when confronted with fog, nighttime illumination, or other adverse visual conditions that frequently occur in inland waterways.

Image dehazing appears to be a direct solution for mitigating visual degradation. However, most dehazing methods are optimized for visual restoration rather than detection-oriented feature preservation. Recent models such as DEA-Net emphasize texture enhancement and global contrast recovery⁷. In inland waterway scenes characterized by strong surface reflections and backlighting, such enhancement may distort vessel contours or introduce artifacts that negatively affect detection. Moreover, a simple cascade of dehazing and detection increases computational cost and inference latency, which conflicts with the real-time requirements of edge-based surveillance systems. Several studies have explored integrated dehazing-based ship detection, either through joint optimization strategies⁸ or two-stage pipelines that combine image enhancement with modified detectors^9,10. These approaches typically rely on heavy architectures or unconditional enhancement mechanisms, which limits their adaptability and efficiency in continuous monitoring scenarios.

Another major challenge lies in the lack of inland waterway-specific datasets captured from ground-level visual sensors. Existing ship detection datasets often exhibit limited environmental diversity^1,11,12, or are constructed from aerial and satellite imagery, such as HRSC2016¹³, FGSD¹⁴, and SSDD¹⁵. These datasets fail to reflect the visual constraints of fixed inland surveillance systems. As a result, models trained on such data tend to generalize poorly when deployed in real inland environments affected by fog, rain, or water-surface reflections¹⁶.

To address these challenges, we propose RT-FogNet, a real-time ship detection framework designed for low-visibility inland waterway environments. RT-FogNet integrates visual enhancement and object detection into a compact end-to-end architecture, enabling robust performance with efficient inference on edge devices. The detection backbone is built upon an enhanced YOLOv10 architecture, in which a Dilated Shared Spatial Pyramid Pooling Fast module is introduced to improve multi-scale feature extraction under complex visual conditions. In addition, we design a lightweight Water Surface Image Dehazing module that adaptively restores image clarity through a domain-aware compression strategy. To support comprehensive evaluation, we construct the Inland Waterway Ship Dataset, a real-world benchmark containing 9,305 annotated images and 16,103 ship instances collected under diverse environmental conditions. This work aims to bridge the gap between academic ship detection research and practical inland monitoring applications.

The main contributions of this work are summarized as follows:

1. RT-FogNet framework: A deployable real-time ship detection system that integrates visual enhancement and object recognition to support inland monitoring under degraded visibility conditions.

2. Backbone-level structural optimization: An enhanced YOLOv10-based detector equipped with a Dilated Shared Spatial Pyramid Pooling Fast module, which improves multi-scale feature representation and detection robustness in foggy, backlit, and reflective scenes.

3. Inland Waterway Ship Dataset: A new dataset comprising 9,305 ground-level images and 16,103 annotated ship instances captured from real inland environments. The dataset covers diverse weather conditions, vessel categories, and lighting variations, providing a realistic benchmark for inland ship detection.

Related work

Ship detection based on object detection

Ship detection can be regarded as a specialized branch of object detection, yet it faces distinct challenges arising from vessel diversity and complex aquatic environments^2,3. Consequently, most modern ship detection approaches are developed on top of general object detection frameworks, particularly convolutional neural network (CNN) and transformer-based architectures.

Early object detection methods are predominantly CNN-based and aim to balance detection accuracy and computational efficiency. The R-CNN family has established region-based detection paradigms^17–19, while one-stage detectors such as YOLOv1 reformulate detection as a regression problem and enable real-time inference²⁰. Subsequent models, including SSD²¹ and RetinaNet²², improve small-object detection and alleviate class imbalance through multi-scale feature representations and focal loss. The YOLO series further evolves through anchor optimization and backbone refinement^23–26. Recent versions eliminate Non-Maximum Suppression and improve deployment efficiency^27–30. Lightweight architectures such as MobileNet^31,32 and EfficientDet³³ achieve competitive accuracy with reduced computational cost, making them suitable for edge deployment.

Transformer-based detectors introduce global attention mechanisms into object detection. DETR enables end-to-end training without anchors or Non-Maximum Suppression³⁴, but it suffers from slow convergence and limited performance on small objects. Later works, including Swin Transformer³⁵ and RT-DETR³⁶, improve feature fusion and inference speed. Despite these advances, transformer-based detectors remain computationally expensive and are less frequently adopted in high-resolution, real-time industrial applications. As a result, CNN-based detectors continue to dominate practical deployments due to their favorable trade-off between accuracy and efficiency.

Building upon these detection frameworks, ship detection research progressively shifts from handcrafted features to data-driven deep learning approaches. Early studies rely on image processing techniques and manually designed features. For example, Li et al. construct the SSDD dataset and combine Constant False Alarm Rate algorithms with SAR imagery to suppress surface clutter¹⁵. However, such methods exhibit limited robustness in complex aquatic environments and struggle with small or densely distributed vessels.

To improve robustness under foggy or low-visibility conditions, several studies explore the integration of image enhancement and ship detection. Liu et al. propose a joint optimization framework that couples dehazing and ship detection within a unified model⁸. Song et al. and Ning et al. adopt two-stage pipelines that apply image enhancement prior to YOLO-based detectors^9,10. Although these approaches report performance improvements under degraded conditions, they typically rely on heavy model architectures or unconditional image enhancement. This reliance limits their adaptability and efficiency in continuous, real-time inland surveillance.

With the continued adoption of deep learning, ship detection performance improves significantly. Zhang et al. report notable gains using Faster R-CNN on the HRSC2016 dataset³⁷, while YOLO-based models enable real-time ship detection in both optical and SAR imagery. Several ship-specific adaptations are proposed, including STD-YOLOv5 with attention and receptive field enhancement⁴, DSSM-LightNet with dual-kernel convolution for efficiency⁶, and multi-scale dilated attention mechanisms for arbitrary-oriented ship detection³⁸.

In addition to visible-light-based ship detection, infrared imaging and infrared–visible fusion approaches have demonstrated strong robustness under extremely low-visibility conditions. Recent works have explored deep learning frameworks for infrared ship detection in foggy and nighttime scenarios^39–41, while cross-modal fusion strategies further enhance structural consistency and detection stability in degraded environments⁴². Although these methods provide valuable insights for maritime perception, they generally require dedicated infrared or multimodal sensing systems. In many inland waterway surveillance applications, monitoring systems are primarily equipped with visible-light cameras, making it practically important to improve detection robustness within the visible spectrum under diverse environmental conditions.

Despite these advances, ship detection in inland waterways remains challenging. Limited data availability hampers fine-grained vessel classification. Complex backgrounds, water-surface reflections, and frequent visibility degradation often lead to false detections and unstable performance^2,3. These limitations highlight the need for detection frameworks that are robust to adverse visual conditions while remaining efficient for real-time, shore-based deployment.

Image dehazing

Image dehazing is a fundamental task in computer vision that aims to restore scene visibility from images degraded by atmospheric scattering⁴³. Over the past decade, research in this field has gradually shifted from physics-based modeling to data-driven learning and multi-task optimization approaches⁴⁴.

Early dehazing methods are primarily based on atmospheric scattering theory and handcrafted priors. The Dark Channel Prior (DCP) estimates transmission maps by assuming that at least one color channel contains near-zero intensity in haze-free regions⁴⁵. Subsequent approaches, such as the Color Attenuation Prior, improve adaptability by modeling depth-dependent color attenuation⁴⁶. However, these methods rely heavily on manually designed features and often suffer from halo artifacts and limited generalization in complex scenes.

With the advancement of deep learning, convolutional neural network based methods have achieved substantial progress in image dehazing. Early CNN models, including DehazeNet⁴⁷ and AOD-Net⁴⁸, adopt end-to-end architectures to directly estimate haze-free images or transmission maps, significantly improving both accuracy and inference efficiency. More recent architectures, such as MSBDN⁴⁹, exploit multi-scale feature fusion to enhance detail recovery. DEA-Net further improves visual restoration by introducing detail-enhanced convolutions and content-guided attention mechanisms, enabling more effective texture enhancement and global contrast recovery⁷.

Beyond supervised learning, recent studies explore unsupervised and contrastive learning strategies to reduce reliance on paired training data. Cycle-Dehaze employs cycle-consistency loss to enable dehazing without ground-truth images⁵⁰. FSAD-Net introduces fog-adaptive modulation to handle varying haze densities⁵¹, while CLDN constrains latent-space representations to mitigate color distortion during restoration⁵².

Despite these advances, many learning-based dehazing methods exhibit limited generalization when applied to real-world scenes that differ from training distributions. To alleviate this issue, domain adaptation techniques have been proposed. SAUD enhances robustness by disentangling depth and fog-related features⁵³, whereas CoA improves synthetic-to-real adaptation through a divide-and-conquer strategy that decomposes the dehazing task into simpler subproblems⁵⁴.

Ship detection datasets

Public datasets have played an essential role in advancing ship detection research, yet their sensing modalities and scene distributions vary considerably. In optical remote sensing, HRSC2016 provides high-resolution satellite imagery with detailed annotations and has been widely adopted for ship detection and recognition¹³. ShipRSImageNet further extends optical benchmarks by offering a large-scale, fine-grained dataset that covers multiple sensors and geographic regions⁵⁵. For SAR-based ship detection, SSDD has become a representative benchmark and includes ships captured by different SAR sensors under varying resolutions and sea conditions¹⁵. More recently, LS-SSDD has been released to support large-scale ship detection in wide-area SAR scenes⁵⁶. Beyond remote sensing, shore-based vision datasets have also been explored. SeaShips collects ship images from real-world coastal surveillance videos and provides category annotations for common vessel types¹¹. In addition, MCMOD includes multiple maritime object categories and supports general maritime detection evaluation⁵⁷.

Despite their contributions, most existing ship detection datasets are collected from satellite or aerial platforms and primarily focus on open-sea or coastal environments^13,15,55,56. These viewpoints differ substantially from fixed, ground-level cameras commonly deployed along inland waterways. As a result, a pronounced domain gap arises in scale, perspective, and background composition, which limits the direct transferability of models trained on remote sensing data to inland surveillance scenarios.

Shore-based datasets such as SeaShips and MCMOD partially alleviate the viewpoint mismatch between aerial maritime datasets and practical surveillance systems; however, they are primarily collected in coastal regions and provide limited coverage of confined inland waterways, dense vessel traffic, and complex man-made riverbank backgrounds^11,57. Recent efforts have extended maritime benchmarks toward inland and adverse environmental conditions. For example, the MEIWVD dataset⁵⁸ investigates inland waterway object detection under multiple environmental settings, and the MAMSD dataset⁴¹ focuses on ship detection in foggy maritime scenes with structural landmark annotations. These works contribute valuable resources for evaluating environmental robustness in maritime detection systems. Nevertheless, comprehensively modeling the visual complexity of shore-based inland surveillance-particularly under low-visibility conditions such as dense fog, backlighting, water-surface reflections, and nighttime illumination-remains an important and ongoing challenge in current research.

Method

Overview

RT-FogNet is developed to enable robust and real-time ship detection in inland waterways under diverse visibility conditions, including dense fog, glare, and surface reflections. As illustrated in Figure 1, the framework employs a conditional two-stage pipeline to process raw input from fixed surveillance cameras.

Fig. 1.

Overview of the proposed RT-FogNet framework.

First, the system evaluates the input image quality using a lightweight visibility assessment mechanism. If the image is captured under low-visibility conditions, it is routed through the Water Surface Image Dehazing (WSID) module, which restores contrast and structural details via a domain-adaptive dehazing strategy. If the visibility is sufficient, WSID is bypassed and the image is forwarded directly to the detection module. This conditional design reduces latency and power consumption in normal conditions while ensuring robust preprocessing under adverse weather.

In both cases, detection is performed using an enhanced YOLOv10-based backbone integrated with a Dilated Shared Spatial Pyramid Pooling Fast (DS-SPPF) module. This design improves multi-scale feature representation while maintaining computational efficiency, enabling accurate and stable detection in diverse inland waterway scenarios. The integration of adaptive preprocessing and an optimized detection backbone forms a unified end-to-end architecture deployable on edge devices with minimal resource overhead.

Water surface image dehazing module

The WSID module is an enhanced front-end dehazing unit built upon the CoA algorithm⁵⁴, specifically redesigned for integration into real-time ship detection systems. As illustrated in Figure 2, WSID operates through a two-stage collaborative process: Model Compression (MoC) in the synthetic domain and Bilevel Adaptation (BiA) in the real domain.

Fig. 2.

Workflow of the WSID Module.

In the MoC stage, a teacher–student distillation scheme is applied to synthetic foggy images, producing a lightweight student network with reduced parameter complexity. In the BiA stage, cross-domain adaptation is achieved via hierarchical optimization: low-level layers are supervised with real-domain foggy data, while high-level representations are stabilized using an Exponential Moving Average (EMA) strategy.

The WSID module adopts a teacher–student architecture. The teacher network uses an input projection of channels, followed by multi-scale branches with channel dimensions and a fusion layer reducing features to 16 channels before the final reconstruction. The lightweight student employs a reduced configuration of input channels and branches , followed by fusion to 8 channels and an output layer.

In the MoC stage, the student is optimized on synthetic data using

1

where and denote teacher and student outputs, respectively. In the BiA stage, real-domain adaptation is performed via

2

while the teacher parameters are updated using exponential moving average .

Compared to the original CoA framework, WSID does not aim to introduce a new dehazing algorithm. Instead, it adapts CoA as a lightweight, detection-oriented front-end module for real-time inland waterway surveillance. The key difference lies in its adaptive integration strategy: WSID is designed to be selectively coupled with the detection backbone according to scene characteristics and deployment requirements, rather than uniformly applied to all inputs. Integrated into RT-FogNet, WSID focuses on preserving structural cues that are critical for ship detection, such as vessel boundaries and contours, while suppressing water-surface interference caused by fog. This design validates the feasibility of coupling dehazing and detection in a unified framework without compromising real-time performance.

Ship detection system

The detection component of RT-FogNet is based on an optimized YOLOv10 architecture. As shown in Figure 3, the YOLOv10 backbone retains the CSPDarkNet-53 structure while introducing enhancements such as C2fCIB and PSA modules to expand the receptive field and strengthen feature representation. The traditional Spatial Pyramid Pooling Fast (SPPF) module is replaced with the proposed DS-SPPF module, which fuses fine-grained and contextual features to improve semantic extraction under fog interference. Additionally, a hybrid PANet–FPN neck enables cross-scale feature aggregation, facilitating the detection of vessels with diverse shapes and visibility levels. The detection head adopts an anchor-free design to simplify training and improve inference efficiency.

Fig. 3.

Architecture of the ship detection system.

A distinctive feature of YOLOv10 is the removal of Non-Maximum Suppression (NMS) during inference. Instead, a dual-assignment strategy is used: one-to-many matching during training provides richer supervision, while one-to-one matching during inference avoids redundancy and reduces latency.

Through lightweight design, multi-scale representation, and end-to-end optimization, the YOLOv10-based backbone provides a strong foundation for real-time deployment in low-visibility inland waterway environments.

Dilated shared spatial pyramid pooling fast module (DS-SPPF)

In the original YOLOv10, the Spatial Pyramid Pooling Fast (SPPF) module enlarges the receptive field through parallel max-pooling operations. Although pooling increases contextual awareness, it inevitably discards spatial details, which is undesirable in foggy inland waterway scenarios where vessel contours are often weak and partially occluded. The loss of fine-grained structure may directly affect localization accuracy and small-object recognition.

To address this limitation, we propose the Dilated Shared Spatial Pyramid Pooling Fast (DS-SPPF) module, which replaces pooling with stacked dilated convolutions to preserve spatial resolution while progressively enlarging the receptive field. All convolutions in DS-SPPF adopt a fixed kernel, and multi-scale perception is achieved by varying the dilation rate rather than increasing kernel size. This design enables receptive field expansion without additional downsampling operations, thereby retaining fine-grained spatial structures and subtle contour information that are critical for accurate vessel boundary localization under fog interference and complex background clutter.

As illustrated in Figure 4, the input feature map , where C, H, and W denote the channel number, height, and width, is first compressed by a convolution to reduce channel redundancy. The compressed feature is then sequentially processed by three dilated convolution layers with dilation rates . Finally, the transformed feature is concatenated with the identity branch and fused through a convolution to produce the output feature map.

Fig. 4.

Architecture of the proposed DS-SPPF module.

For a single dilated convolution with dilation rate r, the effective receptive field size is

3

where k is the kernel size (here ). Substituting gives

4

Thus, dilation rates correspond to effective receptive field sizes of 3, 7, and 11, respectively. For stacked dilated convolutions with stride , the receptive field expands cumulatively as

5

where denotes the receptive field after the l-th layer and denotes the dilation rate of that layer. This sequential structure ensures gradual and continuous receptive field growth without sacrificing spatial resolution.

Dilated convolutions may suffer from the gridding effect when dilation rates are improperly configured, resulting in periodic sampling gaps and incomplete spatial coverage. To mitigate this issue, the dilation configuration in DS-SPPF is designed following the theoretical insights of Hybrid Dilated Convolution (HDC)⁵⁹. Specifically, three principles are implicitly satisfied. First, the dilation rates are arranged in a non-periodic pattern rather than simple multiples of a base value, preventing regular alignment of sampling grids. Second, the dilation rates are pairwise coprime, i.e., for , which avoids repeated overlap of sparse sampling positions across layers. Third, the maximum distance between adjacent sampled positions does not exceed the kernel size, ensuring dense spatial coverage without leaving large unsampled regions.

With the dilation set , these conditions are satisfied simultaneously: the sampling grids interleave across layers, receptive field expansion follows a smooth sawtooth-like progression rather than exponential sparsification, and the effective coverage remains spatially continuous. As a result, DS-SPPF avoids the checkerboard artifacts typically associated with dilated convolutions while preserving fine structural information that is crucial for vessel boundary integrity under fog interference.

To further control computational overhead, DS-SPPF adopts a weight-sharing strategy across the stacked dilated layers. Let and denote the input and output channels, and the kernel size. If independent kernels are used for N dilated layers, the parameter count is

6

With shared weights, the parameter count becomes

7

yielding a reduction ratio of

8

Compared with existing multi-scale aggregation modules, DS-SPPF follows a lightweight and deployment-oriented design philosophy. The original SPPF²⁶ enlarges the receptive field via repeated max-pooling operations, which enhances contextual perception but inevitably discards fine-grained spatial details due to its non-learnable nature. In contrast, DS-SPPF replaces pooling with stacked dilated convolutions, enabling learnable receptive field expansion while preserving spatial resolution. Compared with Atrous Spatial Pyramid Pooling (ASPP)⁶⁰, which employs parallel atrous convolutions with independent parameters at different dilation rates, DS-SPPF adopts a sequential dilation structure with shared weights, thereby reducing parameter redundancy and maintaining smoother feature continuity across dilation stages. Relative to Receptive Field Block (RFB)⁶¹, which introduces multiple heterogeneous branches to approximate complex receptive field patterns, DS-SPPF progressively enlarges the receptive field in a compact manner without introducing additional branch complexity. This design balances contextual aggregation and computational efficiency, making it particularly suitable for real-time inland waterway monitoring scenarios.

Therefore, the proposed DS-SPPF module enlarges the receptive field through structured dilation while maintaining dense spatial sampling, preserves spatial resolution by avoiding pooling-induced information loss, and significantly reduces parameter redundancy through weight sharing. These properties collectively enhance multi-scale feature aggregation and structural continuity, making DS-SPPF particularly suitable for real-time inland ship detection in low-visibility environments.

The inland waterway ship dataset (IWSD)

Dataset construction and characteristics

To overcome the limitations of existing ship detection datasets-such as limited category diversity, insufficient sample sizes, and inadequate coverage of adverse weather conditions-we construct a high-quality, real-world image dataset named the Inland Waterway Ship Dataset (IWSD). This dataset is designed to support comprehensive evaluation of the proposed RT-FogNet framework under complex visual conditions.

The IWSD dataset comprises 9,305 high-resolution images collected from representative inland waterways in China, including high-traffic routes such as the Yangtze River and the Wulong River. In total, the dataset contains 16,103 manually annotated instances. All images are captured using fixed-position shore-based surveillance cameras, ensuring consistent viewpoints and stable imaging conditions across multiple seasons and times of day. IWSD covers nine representative categories of inland vessels, namely cargo ships, buoy tenders, law enforcement vessels, tankers, tugboats, fishing boats, ferries, engineering ships, and recreational boats, spanning diverse operational functions and exhibiting substantial variation in scale, structure, and visual appearance. As shown in Figure 5, panel (a) illustrates the distribution of samples across vessel categories, while panel (b) presents the dataset split, consisting of 7,499 training images, 906 validation images, and 900 test images.

Fig. 5.

Statistical overview of the IWSD dataset. (a) Distribution of samples across ship categories; (b) Dataset partitioning into training, validation, and test sets.

All samples were annotated in COCO format, with bounding boxes manually labeled through a collaborative process involving experienced waterway surveillance personnel and AI engineers. The annotation pipeline included multiple rounds of quality control to ensure high precision and consistency. Notably, the dataset follows a long-tailed distribution: while common vessel types such as cargo ships and recreational boats are well represented, others-such as law enforcement vessels and ferries-appear less frequently. This class imbalance, visualized in Figure 5, reflects the natural occurrence of ships in real-world waterways and presents meaningful opportunities for research on few-shot learning and class rebalancing techniques.

As summarized in Table 1, the IWSD dataset exhibits a pronounced multi-scale distribution across the training, validation, and test splits. Considering the entire dataset, small-scale instances account for 49.2% of all annotated objects, medium-scale instances for 42.7%, and large-scale instances for only 8.1%. This distribution indicates that nearly half of the vessels in IWSD appear at small spatial scales. Such a pattern is consistent with real inland shore-based surveillance settings, where fixed cameras often capture vessels at long distances, resulting in small projected object sizes in the image plane.

Table 1.

Statistical overview of the IWSD dataset, including image counts and object-scale distribution based on the COCO definition (, –, pixels).

Split	#Images	#Instances	Small ()	Medium (–)	Large ()
Train	7,499	12,762	4,427 (34.7%)	6,930 (54.3%)	1,405 (11.0%)
Val	903	1,600	955 (59.7%)	539 (33.7%)	106 (6.6%)
Test	903	1,741	1,082 (62.1%)	567 (32.6%)	92 (5.3%)
Total	9,305	16,103	6,464 (40.1%)	8,036 (49.9%)	1,603 (10.0%)

A split-wise analysis further reveals that the training set is dominated by medium-scale ships (54.3%), whereas the validation and test sets contain substantially higher proportions of small-scale objects, reaching 59.7% and 62.1%, respectively. In contrast, large-scale vessels remain scarce across all splits, accounting for less than 10% of the total instances. This structural imbalance in object scale increases the difficulty of generalization, as the detector must remain highly sensitive to small objects while preserving stable contextual perception for medium and large vessels. Moreover, the discrepancy in scale distribution between the training and evaluation sets further elevates the robustness requirements for real-world deployment.

Overall, these scale statistics demonstrate that IWSD exhibits not only a long-tailed category distribution but also a pronounced imbalance in object scale. Consequently, small-object detection capability becomes a critical criterion for evaluating the robustness and generalization performance of ship detection models in inland waterway environments.

In terms of environmental diversity, IWSD covers a wide range of real inland surveillance conditions, including clear skies, overcast weather, rain–fog scenarios, strong backlighting with water-surface reflections, and nighttime scenes. As summarized in Table 2, nighttime images account for approximately 34.7% of the dataset (3,231 images), making low-light and illumination-interference conditions a major component of IWSD. Rain–fog scenes further contribute 19.1%, while clear and overcast conditions account for 16.1% and 17.2%, respectively. This distribution highlights that more than half of the dataset is collected under visually challenging environments, rather than ideal clear-weather conditions.

Table 2.

Scene statistics of the IWSD dataset.

Scene Type	#Images	Ratio (%)
Clear (Good visibility)	1500	16.1
Overcast (Low contrast)	1600	17.2
Rain-Fog (Rainy and foggy conditions)	1774	19.1
Backlight/Reflection (Glare)	700	7.5
Night (Low-light/illumination interference)	3231	34.7
Other (Rare/uncategorized conditions)	500	5.4
Total	9305	100.0

As illustrated in Figure 6, the test set includes a substantial number of adverse-environment samples characterized by low contrast, glare interference, haze scattering, and nighttime illumination variations. Such scene diversity reflects the intrinsic complexity of real-world shore-based monitoring systems and enables comprehensive evaluation of detection algorithms under practical deployment conditions. By incorporating a high proportion of low-visibility and interference-prone scenes, IWSD provides a rigorous benchmark for assessing robustness and generalization capability in inland waterway ship detection.

Fig. 6.

Sample images from the IWSD test set. (a-c): scenes under dense fog, overcast, and rainy conditions. (d-f): images affected by nighttime illumination interference and low-light visibility.

IWSD is a publicly available dataset designed for multi-class ship detection in real-world inland waterway environments under diverse and often degraded visibility conditions. It provides a realistic and challenging benchmark for training and evaluating deep learning–based ship detection models, particularly in scenarios affected by dense fog, nighttime illumination interference, and low-light weather. By emphasizing low-visibility and visually complex scenes, IWSD facilitates the development and systematic evaluation of anti-interference detection algorithms tailored to practical shore-based surveillance systems.

Comparison with existing ship detection datasets

To further clarify the characteristics and practical advantages of IWSD, we compare it with representative ship detection datasets commonly used in existing studies, including HRSC2016¹³, SSDD¹⁵, LS-SSDD⁵⁶, ShipRSImageNet⁵⁵, SeaShips¹¹, and MCMOD⁵⁷. The comparison focuses on sensing modality, acquisition platform, annotation type, scene coverage, and suitability for inland waterway surveillance.

As summarized in Table 3, most existing datasets are collected from satellite or aerial platforms and primarily target open-sea or coastal environments. Although such datasets provide large-scale coverage and diverse viewpoints, they differ substantially from fixed, ground-level cameras commonly deployed along inland waterways. This discrepancy introduces a notable domain gap in perspective, object scale, and background complexity, which often degrades model performance when transferred to inland surveillance scenarios.

Table 3.

Comparison between IWSD and representative ship detection datasets (k denotes thousands).

Dataset	Modality	Platform	#Images	#Classes	#Instances	Scenes
HRSC201613	Optical	Satellite	1.1k	28	2.9k	Open sea, ports
SSDD15	SAR	Satellite	1.2k	1	2.3k	Open sea, coastal
LS-SSDD56	SAR	Satellite	9k	15	6k	Wide-area sea
ShipRSImageNet55	Optical	Satellite/Aerial	3.5k	50	17k	Diverse maritime
MCMOD57	Optical	Shore-based	16k	10	98k	Mixed maritime
SeaShips11	Optical	Shore-based	7k	6	8k	Coastal surveillance
IWSD (Ours)	Optical	Shore-based	9.3k	9	16k	Inland, low-vis.

Shore-based datasets such as SeaShips and MCMOD partially reduce this gap by adopting ground-level viewpoints. However, they are mainly captured in coastal regions and rarely include confined waterways, dense vessel traffic, or complex man-made backgrounds that frequently appear in inland environments. In addition, adverse visual conditions that are critical for inland monitoring, including dense fog, nighttime illumination interference, and low-light scenes, are either underrepresented or not explicitly annotated in existing benchmarks.

In contrast, IWSD is specifically designed for inland waterway surveillance. All images are captured using fixed-position shore-based cameras deployed along representative inland rivers. The dataset emphasizes low-visibility conditions and includes diverse weather and illumination scenarios encountered in real-world operations. Furthermore, IWSD provides multi-class annotations for inland vessel types, enabling fine-grained evaluation of detection robustness under challenging visual conditions. These properties make IWSD a complementary and necessary benchmark to existing datasets, particularly for the development and evaluation of real-time, anti-interference ship detection systems.

Experiments

To verify the detection performance of the proposed RT-FogNet framework in complex inland waterway environments, we compare it with multiple state-of-the-art (SOTA) object detectors across three datasets: MS COCO⁶², SeaShips¹¹, and the IWSD. All models are trained under identical hardware environments and preprocessing pipelines, with hyperparameters tuned according to the best practice for each method.

Experimental Setup

All experiments are implemented based on the YOLOv10 architecture to ensure consistent comparison. The batch size is set to 256, balancing GPU memory constraints and gradient stability. Optimization is performed using Stochastic Gradient Descent (SGD) with a momentum of 0.937 and weight decay of for L2 regularization, mitigating overfitting and improving generalization. The IWSD module is trained using the Adam optimizer to update the parameters of all submodules. The hyperparameters , , and are set to 0.9, 0.999, and , respectively. The initial learning rate is set to and is gradually decayed to following a cosine annealing schedule.

Experiments are conducted on a high-performance computing platform equipped with two NVIDIA H800 GPUs (80 GB VRAM each), running Ubuntu 22.04.3 LTS with CUDA 12.4. PyTorch 2.0 serves as the primary deep learning framework. To ensure reproducibility, the random seed is fixed at 42, and all experiments are executed within a Conda-managed virtual environment with fixed library versions.

Unless otherwise specified, network hyperparameters (e.g., anchor sizes, pyramid levels) and data preprocessing strategies (e.g., HSV augmentation, affine transformations) follow the baseline configuration. Each experiment is repeated three times under identical settings, and mean values are reported to minimize variance caused by randomness.

Evaluation metrics

Detection performance is measured using , following the COCO evaluation protocol. For convenience, is referred to as throughout this paper.

AP is computed as the area under the precision–recall (P–R) curve:

9

where TP, FP, and FN represent true positives, false positives, and false negatives, respectively.

The interpolated AP is:

10

where denotes the number of recall sampling points in the COCO setting. averages AP across 10 IoU thresholds from 0.50 to 0.95 (step 0.05):

11

We also report (AP at IoU=0.50) for supplementary insight.

Dataset overview

Experiments are conducted on three datasets to evaluate adaptability across general and inland low-visibility environments:

Microsoft Common Objects in Context (MS COCO) is the most authoritative and widely adopted benchmark in the general object detection field. It contains over 200,000 annotated images spanning 80 object categories, with diverse scenarios involving occlusions, lighting variations, and cluttered backgrounds. Although COCO is not ship-specific, its scale, diversity, and established evaluation protocol make it an indispensable reference for assessing generalization capability in complex visual environments.

SeaShips is a domain-specific dataset for inland waterway vessel detection, consisting of approximately 7,000 images captured under various operational conditions in ports and rivers. Compared to COCO, it provides richer context for inland scenarios, including water-surface reflections, piers, and dense traffic scenes. However, most images are captured in high-visibility environments, limiting its coverage of severe low-visibility conditions.

The Inland Waterway Ship Dataset (IWSD), introduced in this work, contains 9,305 high-resolution images collected from major inland waterways in China, such as the Yangtze River and the Wulong River, using fixed shore-based surveillance cameras. The dataset covers diverse weather and lighting conditions, including clear, overcast, rainy, nighttime, and dense fog scenarios, which reflect realistic challenges encountered in practical monitoring systems. IWSD provides 16,103 annotated ship instances across nine vessel categories, namely cargo ships, law enforcement vessels, buoy tenders, tankers, tugboats, fishing boats, ferries, engineering ships, and recreational vessels. All annotations follow the COCO format, ensuring compatibility with mainstream object detection frameworks and facilitating transfer learning across different models and tasks.

Comparative experiment

To comprehensively evaluate the detection performance of the proposed RT-FogNet framework in inland waterway scenarios, we compare it against multiple state-of-the-art (SOTA) object detection models across three datasets: the general-purpose MS COCO⁶², the inland waterway-focused SeaShips¹¹, and the newly constructed Inland Waterway Ship Dataset (IWSD). All models are trained under identical hardware conditions, with consistent preprocessing strategies. Hyperparameters for each model are set to the best configurations reported in their respective publications. The primary evaluation metric is , denoted as in this paper.

Evaluation on the COCO Dataset

MS COCO is the most authoritative benchmark for general object detection, covering 80 categories across complex real-world scenes. To isolate the contribution of the detection backbone, the WSID module is excluded in this experiment, enabling direct comparison with representative anchor-based and anchor-free detectors, including EfficientDet³³, RetinaNet²², Faster R-CNN¹⁹, and YOLOv5 through YOLOv10^{26,27,29,63,64}.

As shown in Table 4, RT-FogNet achieves an of 38.8%, outperforming YOLOv10-N by 1.1 percentage points. Consistent performance gains are observed across all object scales, with of 17.6%, of 39.6%, and of 58.0%. Notably, RT-FogNet maintains the same parameter count of 2.8M and a comparable inference speed to the baseline, achieving 508 FPS compared with 542 FPS for YOLOv10-N. These results demonstrate that the proposed DS-SPPF module effectively enhances multi-scale receptive field aggregation without introducing additional computational overhead. The favorable balance between accuracy and efficiency further indicates strong generalization capability across diverse detection scenarios.

Table 4.

Comparative results on the COCO 2017 dataset. All AP-related metrics are in percentage (%).

Method	Size						FPS	Params	FLOPs
EfficientDet33	512	34.3	54.2	12.0	37.3	50.2	95	3.9 M	2.5 G
RetinaNet-5022	800	35.7	55.0	18.9	38.9	46.3	14	29 M	165 G
Faster R-CNN19	800	27.2	48.4	6.6	28.6	45.0	8	42 M	180 G
YOLOv5-N26	640	28.0	45.7	14.0	31.8	36.6	735	1.9 M	4.5 G
YOLOv6-N63	640	37.0	52.4	16.8	40.2	52.6	1187	4.7 M	11.4 G
YOLOv7-Tiny64	416	33.3	-	-	-	-	1196	6.2 M	5.8 G
YOLOvX-N65	416	32.8	50.3	14.0	35.5	48.3	1143	5.1 M	6.5 G
YOLOv8-N27	640	36.9	52.6	15.3	35.6	54.7	734	3.2 M	8.7 G
YOLOv10-N29	640	37.7	53.2	17.2	37.8	54.9	542	2.8 M	8.7 G
RT-FogNet (Ours)	640	38.8	54.6	17.6	40.8	58.0	508	2.8 M	8.4 G

Evaluation on the SeaShips dataset

The SeaShips dataset targets inland and near-coastal ship detection under standard visibility conditions. The comparison involves lightweight generic detectors, including MobileNet v2⁶⁶ and STD-YOLO⁴, ship-specific detection models such as DSSM-LightNet⁶ and YOLOSeaShip⁶⁷, a cascade-based dehazing and detection approach based on DEA-Net and YOLOv10-N^7,29, as well as the proposed RT-FogNet variants.

As shown in Table 5, lightweight detectors like MobileNet v2 deliver high inference speed but exhibit limited detection accuracy. Ship-specific methods, including DSSM-LightNet and YOLOSeaShip, improve detection performance by leveraging vessel-oriented structural features, although trade-offs between accuracy and computational efficiency remain.

Table 5.

Comparative results on the SeaShips dataset. All AP-related metrics are in percentage (%).

Method			Precision	Recall	FPS	Params	FLOPs
MobileNet v266	76.0	97.7	96.3	94.4	268	3.8 M	10 G
STD-YOLO4	61.2	95.7	89.6	90.5	258	16.3 M	35.7 G
DSSM-LightNet6	78.2	98.5	97.7	94.8	469	2.8 M	7.1 G
YOLOSeaShip67	77.5	97.6	94.8	95.9	366	4.8 M	11 G
YOLOv10-N29	81.1	98.5	97.4	96.5	542	2.8 M	8.7 G
DEA-Net+YOLOv10-N7,29	81.4	98.1	96.9	96.9	248	6.7 M	38.4 G
RT-FogNet (Ours)	82.4	98.9	97.8	97.9	508	2.8 M	8.4 G
RT-FogNet+ (Ours)	82.6	98.8	97.9	98.1	223	5.2 M	12.5 G

RT-FogNet attains an of 82.4% on the SeaShips dataset, accompanied by an of 98.9%, a precision of 97.8%, and a recall of 97.9%. The model maintains real-time inference with an inference speed of 508 FPS and a compact parameter size of 2.8M, indicating robust performance under the normal visual conditions typical of shore-based surveillance. RT-FogNet+ further enhances detection accuracy, reaching an of 82.6% and achieving the highest recall of 98.1% among all evaluated methods. The performance gain remains moderate, which aligns with the characteristics of the SeaShips dataset, where low-visibility scenes are relatively scarce and most samples are captured under clear or mildly degraded conditions.

By comparison, the cascade-based DEA-Net and YOLOv10-N approach achieves an of 81.4%, but introduces a substantial increase in model complexity and a pronounced reduction in inference speed. This observation indicates that directly integrating generic dehazing networks into detection pipelines can incur unnecessary computational overhead when applied to datasets dominated by standard visibility conditions.

Overall, the SeaShips experiments demonstrate that RT-FogNet provides a favorable balance between detection accuracy and efficiency, while RT-FogNet+ offers additional robustness when visual enhancement becomes beneficial. These results further emphasize the importance of tailoring dehazing strategies to dataset characteristics and deployment environments.

Evaluation on the IWSD dataset

The IWSD dataset captures challenging inland waterway scenarios characterized by dense fog, strong backlighting, nighttime illumination, and water-surface reflections. To evaluate robustness under these degraded visual conditions, we compare RT-FogNet without WSID and RT-FogNet+ with WSID against representative baseline methods.

As reported in Table 6, RT-FogNet without WSID achieves an of 40.8%, outperforming YOLOv10-N (38.7%) while maintaining high inference efficiency at 508 FPS. This improvement is consistently reflected across object scales. Compared with YOLOv10-N, RT-FogNet increases performance from 29.4% to 31.7% on small targets, from 46.8% to 48.6% on medium targets, and from 61.3% to 65.8% on large targets. These results indicate that the DS-SPPF module effectively enhances multi-scale feature aggregation, with particularly notable gains on large and medium objects where contextual modeling is crucial, while still providing stable improvements for small-scale vessels under visually cluttered conditions.

Table 6.

Performance on the IWSD dataset under challenging visual conditions. All AP metrics in %.

Method			APS	APM	APL	Precision	Recall	FPS	Params	FLOPs
MobileNet v266	34.1	43.7	25.6	40.7	58.3	69.2	37.5	268	3.8 M	10 G
STD-YOLO4	35.7	44.5	27.4	42.5	60.1	69.2	38.4	258	16.3 M	35.7 G
DSSM-LightNet6	37.2	46.3	29.1	44.6	62.8	72.4	39.7	469	2.8 M	7.1 G
YOLOSeaShip67	34.8	43.8	26.8	41.9	59.4	67.4	37.9	366	4.8 M	11 G
YOLOv10-N29	38.7	48.9	29.4	46.8	61.3	74.5	37.3	539	2.8 M	8.7 G
DEA-Net+YOLOv10-N7,29	46.5	57.1	35.9	54.2	66.5	79.5	48.9	248	6.7 M	38.4 G
RT-FogNet (Ours)	40.8	51.8	31.7	48.6	65.8	78.5	40.6	508	2.8 M	8.4 G
RT-FogNet+ (Ours)	53.9	65.6	44.6	61.7	71.9	88.5	55.6	223	5.2 M	12.5 G

With the WSID module enabled, RT-FogNet+ delivers a substantial performance improvement, reaching an of 53.9% and an of 65.6%. Precision increases to 88.5%, and recall improves markedly to 55.6%, highlighting a significantly reduced rate of missed detections. The scale-specific results further demonstrate the advantage of the integrated framework. On small targets, AP rises sharply to 44.6%, representing a 15.2 percentage point improvement over YOLOv10-N and a clear margin over the cascade-based DEA-Net+YOLOv10 (35.9%). Similar improvements are observed for medium targets (61.7%) and large targets (71.9%), confirming that the proposed joint enhancement and detection strategy benefits objects across all scales.

Although the inference speed decreases to 223 FPS due to the additional computation introduced by the WSID module, the model still satisfies real-time or near-real-time deployment requirements for inland monitoring systems. Overall, DS-SPPF enhances structural feature representation at the feature-extraction level, while WSID further improves small-object sensitivity and robustness under low-contrast and degraded visibility conditions. The combination of these two components enables stable multi-scale performance gains in complex inland waterway environments.

To further analyze the impact of the long-tailed distribution in IWSD, per-class AP results are reported in Table 7. This evaluation provides a more detailed understanding of model robustness across both dominant and minority vessel categories.

Table 7.

Per-class detection performance on the IWSD dataset under challenging visual conditions. All metrics are in percentage (%). Class abbreviations: Fish.=Fishing boats, Ferr.=Ferries, Buoy.=Buoy tenders, Cargo=Cargo ships, Eng.=Engineering ships, Tug=Tugboats, Tank=Oil tankers, Rec.=Recreational boats, Law=Law enforcement vessels.

Method	Fish.	Ferr.	Buoy.	Cargo	Eng.	Tug	Tank	Rec.	Law
YOLOv10-N29	35.0	31.0	43.0	52.0	33.0	45.0	48.0	42.0	19.0	38.7
DEA-Net+YOLOv10-N7,29	44.0	39.0	50.0	59.0	41.0	53.0	55.0	49.0	29.0	46.5
RT-FogNet (Ours)	37.0	33.0	45.0	54.0	35.0	47.0	50.0	44.0	22.0	40.8
RT-FogNet+ (Ours)	52.0	47.0	56.0	64.0	49.0	58.0	60.0	55.0	44.0	53.9

YOLOv10 shows relatively lower performance on tail classes such as Law enforcement vessels and Ferries, which is consistent with the class imbalance in IWSD. The cascade-based DEA-Net+YOLOv10 improves these categories to some extent, indicating that visibility enhancement benefits minority classes; however, the overall gains remain limited, suggesting that simple dehazing–detection coupling does not fully resolve long-tail challenges.

RT-FogNet achieves more stable improvements across most categories, while RT-FogNet+ delivers substantial gains on tail classes-particularly Law enforcement vessels and Engineering ships-without sacrificing performance on head classes such as Cargo ships. These results demonstrate that the proposed integrated framework enhances detection robustness under long-tailed inland waterway scenarios.

The effectiveness of WSID is further illustrated by qualitative results in Figure 7. In the second and third columns, the Oil tanker (YL) targets are only correctly detected by RT-FogNet+, while other methods either miss the targets or produce unstable predictions. These cases correspond to low-contrast and partially occluded scenes, where fine structural cues are easily suppressed by fog and glare. The DS-SPPF module mitigates the gridding effect inherent to dilated convolutions by adopting a dilation configuration guided by the HDC principle, which promotes complementary sampling patterns across layers and avoids periodic coverage gaps. As a result, spatial information is aggregated more continuously, enabling more stable contour representation for minority classes.

Fig. 7.

Qualitative Visualization of Ship Detection Results on the IWSD Dataset. Class abbreviations in the figure denote: HC (Cargo ship), YL (Oil tanker), and HBC (Buoy tender).

Moreover, the improvement on the Oil tanker (YL) category, which belongs to a relatively less frequent class compared with Cargo ships, reflects the model’s robustness under long-tailed distributions. By preserving local details and progressively expanding contextual perception, DS-SPPF enhances feature consistency across scales, while WSID restores visibility-sensitive structural information. The combined effect allows RT-FogNet+ to better generalize to tail classes and challenging visual conditions simultaneously.

Overall, the visualization results are consistent with the quantitative per-class analysis, demonstrating that the integration of DS-SPPF and WSID improves both spatial continuity and class-wise robustness in real-world inland waterway scenarios.

Overall discussion

Across all three datasets, RT-FogNet consistently demonstrates a favorable balance between detection accuracy, model complexity, and inference speed. On the general-domain COCO benchmark, the DS-SPPF-enhanced YOLOv10 backbone achieves superior multi-scale feature representation without increasing parameters or sacrificing real-time throughput. In the SeaShips dataset, which represents typical inland and near-coastal conditions, RT-FogNet preserves fine-grained structural details of vessels, achieving state-of-the-art performance while maintaining sub-2 ms per-image inference time. Most notably, on the challenging IWSD dataset, RT-FogNet+ with the WSID module achieves substantial gains in both and recall, underscoring the advantage of integrating visibility restoration with adaptive feature fusion. These results validate that the proposed framework’s twofold innovation-(1) the DS-SPPF module for receptive field diversification with minimal overhead, and (2) the WSID module for domain-adaptive visual enhancement-enables robust and efficient ship detection across diverse inland waterway visibility conditions.

Ablation study

To systematically evaluate the individual and joint contributions of the core components in the RT-FogNet framework, we conduct ablation experiments on the IWSD dataset, focusing on two primary modules: the DS-SPPF module and the WSID module. The DS-SPPF module is designed to enhance multi-scale feature representation via grouped dilated convolutions with varied dilation rates, while the WSID module acts as a lightweight, domain-adaptive visual enhancement front-end that restores contrast and structural details under fog, glare, and other low-visibility conditions.

Table 8 summarizes the ablation configurations, and Table 9 presents the corresponding performance metrics, including , , precision, recall, inference speed (FPS), and model complexity. In total, seven configurations are tested:

• Baseline: without DS-SPPF and WSID.

• Settings 1–5: baseline plus DS-SPPF variants with different dilation rate combinations. Setting-1 uses (1,2,3), Setting-2 uses (1,2,4), Setting-3 uses (2,3,5), Setting-4 uses (2,4,6), and Setting-5 uses (1,3,5).

• Setting-6: the optimal DS-SPPF configuration (Setting-5) combined with WSID, forming the complete RT-FogNet+.

• Setting-7: WSID applied directly to the baseline without DS-SPPF.

Table 8.

Ablation settings. DS-1 to DS-5 represent different DS-SPPF configurations; WSID denotes the dehazing module.

Setting	DS-1	DS-2	DS-3	DS-4	DS-5	WSID	(%)
Baseline	-	-	-	-	-	-	38.72
Setting-1		-	-	-	-	-	38.94 0.22
Setting-2	-		-	-	-	-	39.21 0.49
Setting-3	-	-		-	-	-	37.88 0.84
Setting-4	-	-	-		-	-	37.99 0.73
Setting-5 (RT-FogNet)	-	-	-	-		-	40.82 2.10
Setting-6 (RT-FogNet+)	-	-	-	-			53.94 15.22
Setting-7	-	-	-	-	-		48.54 9.82

Table 9.

Ablation study results on the IWSD dataset. DS-1 to DS-5 denote different DS-SPPF configurations. All AP-related metrics are reported as percentages (%).

Method			Precision	Recall	FPS	Params	FLOPs
Baseline	38.72	48.86	74.52	37.29	539	2.8M	8.7G
Setting-1	38.94	48.96	74.72	37.39	558	2.7M	8.7G
Setting-2	39.21	49.16	74.81	37.42	556	2.8M	8.7G
Setting-3	37.88	47.84	73.61	36.54	512	2.9M	8.8G
Setting-4	37.99	47.96	73.87	36.96	481	2.9M	8.8G
Setting-5	40.82	51.79	78.47	40.61	508	2.8M	8.4G
Setting-6	53.94	65.55	88.49	55.63	223	5.2M	12.5G
Setting-7	48.54	59.63	80.57	46.28	295	5.3M	12.8G

From the baseline performance of , several key observations emerge. First, among the DS-SPPF variants (Settings 1–5), the (1,3,5) dilation pattern in Setting-5 achieves the best (40.82%) and recall (40.61%), while keeping precision competitive (78.47%) and inference speed high (508 FPS). This confirms that a balanced dilation configuration effectively combines fine-grained detail extraction with long-range context aggregation. Conversely, wider dilation gaps such as (2,3,5) and (2,4,6) in Settings 3 and 4 reduce accuracy, suggesting that overly sparse receptive fields may fail to capture small-scale object features. These findings validate the design principle that balanced multi-scale coverage is crucial for preserving both local resolution and contextual awareness.

Second, the WSID module substantially boosts detection in low-visibility scenes. Integrating WSID with the optimal DS-SPPF (Setting-6) yields a 13.12-point improvement and a 15.02-point recall gain over Setting-5, raising precision to 88.49%. WSID alone (Setting-7) also improves performance by 9.82 points over baseline, though still below the full integration in Setting-6. This highlights the complementary nature of DS-SPPF and WSID: the former enhances multi-scale feature fusion, while the latter improves input quality, and their synergy maximizes robustness.

Finally, the superior performance of RT-FogNet derives from two core innovations: (1) Structural optimization - the DS-SPPF employs weight-shared multi-branch dilated convolutions, expanding the receptive field while reducing parameters to 1/N of an equivalent independent design, enabling efficient deployment without sacrificing speed; (2) Input enhancement - the WSID module uses a compression–bilevel adaptation strategy to maintain stable dehazing across domains, ensuring reliable feature extraction even in low-contrast or degraded inputs.

In summary, the (1,3,5) DS-SPPF configuration combined with WSID (Setting-6) achieves the highest accuracy, recall, and precision while retaining acceptable real-time performance, confirming that structural optimization and visual enhancement together enable RT-FogNet to deliver high accuracy, robustness, and efficiency for inland waterway ship detection.

Conclusion

This paper presents RT-FogNet, a lightweight and robust ship detection framework tailored for inland waterways with low visibility and complex visual interference. The proposed DS-SPPF module enhances multi-scale spatial feature representation through weight-shared dilated convolutions, while the WSID dehazing module restores contrast and structural details under dense fog, strong backlighting, and surface reflections. To support rigorous evaluation, we constructed the Inland Waterway Ship Dataset (IWSD), which encompasses diverse environmental and visibility conditions representative of real-world monitoring scenarios.

Comprehensive experiments on the COCO, SeaShips, and IWSD datasets demonstrate that RT-FogNet achieves a favorable balance between accuracy, computational efficiency, and real-time performance. On the IWSD dataset, the enhanced RT-FogNet+ variant surpasses all baselines by more than 15 percentage points in , highlighting its strong robustness and applicability under adverse visual conditions. The improvements are attributed to the synergy between structural optimization (DS-SPPF) and domain-adaptive visual enhancement (WSID), enabling effective detection across varying scales and degraded inputs.

Nevertheless, integrating WSID introduces additional computational overhead, which may limit deployment on extremely resource-constrained platforms. Furthermore, current evaluations are based on static datasets, and further field trials are necessary to validate long-term stability and adaptability in dynamic operational environments. Future work will explore more lightweight enhancement modules to reduce inference costs, investigate multi-modal sensing strategies-such as integrating radar and infrared-and deploy the system in real-world inland waterway and port surveillance applications.

Author contributions

Conceptualization, S.F., P.Y. and C.X.; methodology, S.F.; software, C.X.; validation, X.T., C.X. and S.F.; formal analysis, C.X., X,T.; investigation, S.F.; resources, P.Y.; data curation, P.Y.; writing—original draft preparation, C.X., S.F.; writing—review and editing, S.F., P.Y.; visualization, C.X.; supervision, S.F.; project administration, S.F, P.Y.; funding acquisition, P.Y. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported in part by the National Key R&D Program of China under Grant No. 2024YFC3015200, the National Natural Science Foundation of China under Grant No. U24A20175, the State Key Laboratory of Bridge Safety and Resilience, and China Merchants Chongqing Communications Technology Research and Design Institute Co. Ltd.

Data availability

The dataset and source code proposed in this study are freely available at https://github.com/Object-Detection-01/RT-FogNet. Both the data and code are openly released to the research community to support academic use and reproducibility. Researchers are encouraged to access, utilize, and build upon these resources for non-commercial scientific purposes. As a condition of use, any published work or public report that uses this dataset must include an explicit citation of this paper. For further details, please refer to the repository.

Declarations

Competing interests

The authors declare no competing interests.