Abstract
Automated Ship License Plate (SLP) recognition is critical for maritime surveillance but faces significant challenges due to non-standard layouts and harsh environmental conditions. This paper introduces SLP34K, the first large-scale, high-diversity dataset for SLP recognition, featuring over 34,000 images captured over four years. Alongside this benchmark, the authors propose a robust baseline model that integrates self-supervised pre-training with a semantic enhancement module. This approach sets a new state-of-the-art for both maritime-specific and general scene text recognition (STR) tasks.
Introduction: The Challenge of Maritime Text Recognition
While automated license plate recognition for automobiles is a mature field, identifying ships remains a formidable task. Ship License Plates (SLPs) are truly "in the wild": they lack standardized fonts, appear in complex multi-line or vertical layouts, and are frequently obscured by environmental factors like fog, rain, or low-light conditions.
Historically, the field has been limited by a "data drought." Existing SLP datasets are typically small (under 6,000 images) and lack the geographic or temporal diversity needed for deep learning models to generalize. This paper addresses these gaps by providing the scale and variety necessary for robust maritime surveillance.
The SLP34K Dataset
The centerpiece of this research is SLP34K, a dataset containing 34,385 images collected from eight different surveillance cameras over a 42-month period. It is 15 times larger than previous datasets and includes detailed manual annotations for text, layout (single-line, multi-line, vertical), and difficulty levels.
Methodology: MAE and Semantic Enhancement
The authors propose a simple yet powerful encoder-decoder baseline designed to handle visual degradation and complex layouts:
- Self-Supervised Pre-training (MAE): Instead of relying on general datasets like ImageNet, the authors use Masked Autoencoders (MAE) to pre-train a Vision Transformer (ViT) encoder directly on the SLP34K dataset. This allows the model to learn maritime-specific visual features and noise patterns.
- Semantic Enhancement Module (SEM): To assist the model in "reading" through blur and noise, a semantic bridge is introduced. During training, a frozen CLIP text encoder provides semantic guidance. A contrastive loss (InfoNCE) aligns visual features with linguistic features, teaching the encoder to infer characters even when they are visually degraded.
- Auto-Regressive Decoder: A standard decoder predicts character sequences from these enhanced visual features.
Key Results and Insights
- Superior Performance: The proposed baseline achieved an accuracy of 83.57% on the SLP34K test set, significantly outperforming established models like PARSeq (70.53%) and ABINet (66.52%).
- Generalization Excellence: Despite its maritime focus, the model achieved state-of-the-art results on general STR benchmarks, including 99.4% accuracy on IIIT5K and 98.8% on ICDAR2013.
- Domain-Specific Gains: Self-supervised pre-training on SLP data proved more effective than standard ImageNet-supervised weights, highlighting the unique textures of maritime environments.
- Semantic Robustness: The SEM provided the most significant gains on "Hard" samples, where visual signals are weakest, by leveraging learned linguistic patterns to correct for noise.
Conclusion and Future Work
SLP34K represents a major step forward for maritime text recognition. While the new baseline handles complex layouts and noise effectively, extreme weather conditions still pose challenges. Future research may focus on end-to-end ship detection and recognition pipelines or the use of temporal video data to further mitigate transient environmental noise.
Summary updated