TL;DR
This paper introduces FSN-YOLO, a novel object detection framework for nearshore vessels that enhances YOLOv8 by integrating a lightweight FasterNet backbone and a Receptive-Field Attention (RFA) mechanism. FSN-YOLO achieves improved accuracy and recall in detecting vessels of various scales, even in challenging maritime conditions, by better capturing fine-grained features and reducing redundant information.
What Problem Does This Paper Address?
Accurate and efficient vessel detection is crucial for maritime safety, transportation, and surveillance. However, detecting vessels in nearshore environments presents significant challenges. These include the vast diversity in vessel types and sizes, the similarity in ship hull shapes, and the disruptive effects of complex environmental conditions like waves and weather. Existing detection methods often struggle to strike a balance between precision, speed, and model size, particularly when dealing with multi-scale targets and occlusions. This can lead to missed detections of smaller vessels or false positives, hindering effective maritime operations.
Methods
The FSN-YOLO framework builds upon the YOLOv8 architecture, introducing several key modifications:
- Lightweight Backbone (FasterNet): The original YOLOv8 backbone is replaced with FasterNet. FasterNet utilizes "partial convolution" (PConv), which processes only a subset of channels, reducing computational load and memory usage while aiming to maintain feature representation quality. This contributes to a balance between speed and model size.
- Receptive-Field Attention (RFA) Mechanism: Integrated into the neck of the network, the RFA mechanism is designed to adaptively focus on salient features within receptive fields. It uses group convolutions and attention weights to dynamically highlight important spatial features and suppress irrelevant information. This helps in capturing fine-grained details and reducing feature redundancy, which is particularly beneficial for detecting small or distant vessels.
- Feature Fusion: The RFA module enhances the multi-scale feature fusion process, allowing the network to better understand complex maritime scenes and improve the discriminative power of ship features.
Datasets
The primary dataset used for training and evaluation in this study is the Seaship7000 dataset. This dataset consists of 7000 high-resolution images (up to 1920x1080 pixels) captured by visual surveillance cameras in the Hengqin New Area, Zhuhai. It features diverse nearshore ship scenarios, including various vessel sizes, lighting conditions, and partial occlusions. The dataset includes six vessel categories: ore carriers (OCs), bulk cargo carriers (BCCs), general cargo ships (GCSs), container ships (CSs), fishing boats (FBs), and passenger ships (PSs). The data was split into training, validation, and test sets in an 8:1:1 ratio.
Key Insights and Ideas
- The Power of Focused Attention: The RFA mechanism demonstrates that intelligently focusing attention on specific spatial features within receptive fields, rather than treating all features uniformly, can significantly boost detection performance, especially in complex environments.
- Balancing Act: Speed vs. Accuracy: By combining a lightweight backbone like FasterNet with an effective attention mechanism, the authors show it's possible to achieve high detection accuracy without a prohibitive increase in model size or computational cost.
- Nearshore Specificity Matters: While general object detection models are improving, domain-specific challenges in nearshore environments (like scale variation and occlusion) require tailored solutions, as evidenced by FSN-YOLO's performance against both general and domain-specific methods.
Key Results
The FSN-YOLO model demonstrated significant improvements over the baseline YOLOv8l on the Seaship7000 dataset:
- Accuracy: Achieved an absolute increase of 0.82% in precision.
- Recall: Achieved an absolute increase of 1.54% in recall.
- [email protected]:0.95: Achieved an absolute increase of 1.56%.
- Overall Performance: The model achieved a precision of 0.989, recall of 0.986, and [email protected] of 0.993.
- Efficiency: While the parameter count increased slightly (to 48.56M), the inference time was maintained at a competitive level (7.4 ms), making it suitable for real-time applications.
- Qualitative Results: Visualizations showed FSN-YOLO's superior ability to detect smaller and more distant vessels, reducing missed detections and false positives compared to baseline models.
Limitations and Future Work
- Parameter Count: While inference time is competitive, the parameter count of FSN-YOLO is higher than the baseline YOLOv8l. The authors suggest that further refinements could potentially reduce this.
- Future Work: The authors suggest that future research could explore further optimizations for parameter reduction and investigate the model's applicability to other maritime surveillance tasks. They also highlight the model's adaptability to dynamic and complex marine environments as a key strength for future development.
Summary updated