Enhanced YOLO v3 Tiny Network for Real-Time Ship Detection From Visual Image

Hao Li, Lianbing Deng, Cheng Yang, Jianbo Liu, Zhaoquan GuIEEE Access 2021 · OpenReview

Abstract

Real-time ship detection from visual images (SDVI) is critical for port management, autonomous navigation, and maritime safety. While standard object detection models exist, they often struggle with the unique challenges of maritime environments, such as small targets and complex background interference. This paper presents an enhanced YOLOv3-Tiny network specifically optimized for SDVI. By implementing custom anchor boxes, replacing max-pooling with convolutional downsampling, and integrating the Convolutional Block Attention Module (CBAM), the proposed model achieves a 97% mAP at 135 FPS. This represents a significant improvement over existing state-of-the-art methods, offering a superior trade-off between real-time performance and detection accuracy.

Introduction

Ship detection is a cornerstone of modern maritime logistics. Traditional methods like Ship Detection from Synthetic Aperture Radar (SDSAR) and Spaceborne Optical Images (SDSOI) are often limited by low resolution and long update cycles. Ship Detection from Visual Image (SDVI) utilizes standard video surveillance, providing high-resolution, real-time data. However, SDVI faces three primary hurdles:

  1. Small Target Detection: Distant ships appear as tiny pixel clusters that are easily lost during standard neural network downsampling.
  2. Environmental Interference: Waves, light reflections, and onshore buildings frequently trigger false positives.
  3. The Speed-Accuracy Trade-off: "Heavy" models like YOLOv3 are accurate but slow, while "tiny" versions are fast but often lack the precision required for complex scenes.

Methodology

The proposed "Enhanced YOLOv3-Tiny" introduces several architectural refinements to the standard lightweight YOLOv3-Tiny framework to address these challenges.

Custom Anchor Boxes

Standard YOLO anchors are designed for a broad range of objects. For maritime surveillance, the authors used K-means clustering on the Seaship dataset to redefine these anchors. The resulting "dumpy" anchors better match the specific silhouettes and aspect ratios of ships, facilitating faster training and higher detection precision.

Comparison of original YOLOv3-Tiny anchors and the redefined "dumpy" anchors optimized for ship detection.

Convolutional Downsampling and Channel Expansion

To preserve features of small targets, standard max-pooling layers were replaced with $3 \times 3$ convolutional layers with a stride of 2. Unlike max-pooling, which discards data, learned convolutions allow the network to compress information while retaining essential features. Furthermore, the prediction channels for the small-target branch were expanded from 128 to 512, deepening the network's ability to extract features from tiny objects.

CBAM Integration

The Convolutional Block Attention Module (CBAM) was integrated into the backbone. CBAM applies both channel and spatial attention, teaching the model to "ignore" the background noise of waves and buildings and focus its computational weight on the structural features of the ships themselves.

Experimental Results

The model was evaluated using the Seaship dataset, containing over 30,000 images of six ship types: ore carriers, bulk cargo carriers, general cargo ships, container ships, fishing boats, and passenger ships.

The enhanced model achieved a 97% mAP, representing a 9.6% improvement over previous saliency-aware CNN benchmarks. In terms of speed, the model runs at 135 FPS on an NVIDIA Titan RTX. While slightly slower than the original YOLOv3-Tiny (188 FPS), it is significantly faster than the full YOLOv3 (70 FPS) while maintaining near-comparable accuracy. Notably, the detection of small targets like fishing boats and passenger ships reached 96.2% and 96.7% AP, respectively.

Conclusion

The Enhanced YOLOv3-Tiny network successfully bridges the gap between high-speed "tiny" models and high-accuracy "heavy" models. By tailoring the architecture to the unique geometry and environmental challenges of maritime scenes, the researchers have provided a robust tool for real-time surveillance. Future work will focus on further optimizing the network for lower-power edge devices where power consumption is a primary constraint.

Summary updated

Abstract

Different from ship detection from synthetic aperture radar (SDSAR) and ship detection from spaceborne optical images (SDSOI), ship detection from visual image (SDVI) has better detection accuracy and real-time performance, which can be widely used in port management, cross-border ship detection, autonomous ship, safe navigation, and other real-time applications. In this paper, we proposed a new SDVI algorithm, named enhanced YOLO v3 tiny network for real-time ship detection. The algorithm can be used in video surveillance to realize the accurate classification and positioning of six types of ships (including ore carrier, bulk cargo carrier, general cargo ship, container ship, fishing boat, and passenger ship) in real-time. Based on the original YOLO v3 tiny network, we have made the following fine tunings. 1) The preset anchors trained on Seaship annotation data have the similar “dumpy” shape as the normal ships, helping the network to achieve faster and better training; 2) Convolution layer instead of max-pooling layer and expanding the channels of prediction network improve the small target detection ability of the algorithm. 3) Due to the problem that large-scale ships are easily disturbed by the onshore building, complex waves and light on the water surface, we introduced attention module named CBAM into the backbone network, which make the model more focused on the target. The detection accuracy of the proposed algorism is obviously better than that of the original YOLO v3 tiny work. Although it is slightly inferior to the Yolo v3 network, it has faster speed than Yolo v3. However, the proposed algorithm is a better trade-off between real-time performance and detection accuracy, and is more suitable for actual scenes. Compared with the SOAT algorithm in Z. Shao et al. (2020), our algorithm has a 9.6% improvement in <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mAP</i> and a faster speed.
BibTeX
@inproceedings{li2021enhanced,
  author = {Hao Li and Lianbing Deng and Cheng Yang and Jianbo Liu and Zhaoquan Gu},
  title = {{Enhanced YOLO v3 Tiny Network for Real-Time Ship Detection From Visual Image}},
  booktitle = {IEEE Access},
  year = {2021},
  url = {https://openreview.net/forum?id=BPuRLPghnVD},
}
Figure