Research map / Label-efficient and robust learning / Self-supervised and contrastive
Towards Ship License Plate Recognition in the Wild: A Large Benchmark and Strong Baseline
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds SLP34K, a 34,385-image ship license plate recognition dataset from a waterway surveillance system, and proposes a baseline combining MAE self-supervised pre-training with CLIP-text contrastive semantic enhancement on a ViT encoder with an autoregressive decoder.
Problem
Ship license plate recognition is limited by the lack of large, public, diverse datasets; existing datasets are small, captured at one location, and unreleased, and SLPs have complex layouts and visual degradation.
Approach
Pre-train a ViT encoder with masked autoencoding (MAE) on task data, then fine-tune with cross-entropy plus an InfoNCE loss aligning visual CLS features to frozen CLIP text features of the plate label, yielding a strong encoder for SLP and scene text recognition.
Key results
- On SLP34K, the proposed baseline with ViT and MAE pre-training plus semantic enhancement reaches 83.57% accuracy, above MAEREC (79.84%) and PARSeq (70.53%) under the same ViT backbone.
- On Union14M-L, the baseline reaches 91.9% versus 88.8% for MAEREC (the previous best on that benchmark).
- Semantic enhancement improves accuracy on every pre-training variant, e.g. SFS pre-training rises from 81.80% to 83.57%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed ViT baseline instead of MAEREC | SLP recognition · SLP34K test | accuracy: 79.84 -> 83.57 | Helped |
| use the proposed ResNet34 baseline instead of ABINet | SLP recognition · SLP34K test | accuracy: 66.52 -> 76.55 | Helped |
| use the proposed method instead of MAEREC | scene text recognition · Union14M-L test | word accuracy: 85.2 -> 87.9 | Helped |
| pre-train with classification on ImageNet instead of training from scratch (ablation) | ViT-B/16 for SLP recognition · SLP34K test | accuracy: 79.28 -> 77.10 | Hurt |
| pre-train with contrastive learning on WebImageText instead of training from scratch (ablation) | ViT-B/16 for SLP recognition · SLP34K test | accuracy: 79.28 -> 57.98 | Hurt |
| pre-train with self-supervised MAE on ImageNet instead of training from scratch (ablation) | ViT-B/16 for SLP recognition · SLP34K test | accuracy: 79.28 -> 80.88 | Helped |
| use ViT instead of ResNet152 (ablation) | SLP recognition · SLP34K test | accuracy: 75.60 -> 79.28 | Helped |
| add semantic enhancement (ablation) | ViT-B/16 pretrained by MAE on SLP34K · SLP34K test | accuracy: 81.80 -> 83.57 | Helped |
| replace ImageNet with SLP34K for MAE pre-training (ablation) | ViT-B/16 for SLP recognition · SLP34K test | accuracy: 80.88 -> 81.80 | Helped |
| add semantic enhancement to the classification-pretrained encoder (ablation) | ViT-B/16 for SLP recognition · SLP34K test | accuracy: 77.10 -> 79.17 | Helped |
| add semantic enhancement to the contrastive-pretrained encoder (ablation) | ViT-B/16 for SLP recognition · SLP34K test | accuracy: 57.98 -> 64.22 | Helped |
| add semantic enhancement to the MAE-pretrained encoder (ablation) | ViT-B/16 pretrained on ImageNet · SLP34K test | accuracy: 80.88 -> 83.57 | Helped |
Surprising findings
- Classification pre-training on ImageNet and CLIP-style contrastive pre-training underperform training from scratch for SLP recognition, while MAE pre-training helps.
Limitations
- Existing SLP methods could not be compared directly because they are not open-source.
- Only one dataset (SLP34K) is used for SLP evaluation, so generalization to other SLP data is untested.
Experimental setup
Datasets: SLP34K Union14M-L ImageNet IIIT5K ICDAR2013 SVT ICDAR2015 SVTP CUTE80
Benchmarks: SLP34K Union14M-L IIIT5K ICDAR2013 SVT ICDAR2015 SVTP CUTE80
Metrics: word accuracy sequence accuracy
Baselines: MATRN SemiMTR ABINet ViTSTR PARSeq MAEREC SVTR SATRN
Builds on: MAE CLIP PARSeq
Closest papers by meaning
- LP-Diff: Towards Improved Restoration of Real-World Degraded License Plate (CVPR 2025 · similarity 0.83)
- Traffic Scene Parsing through the TSP6K Dataset (CVPR 2024 · similarity 0.79)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.79)
- Neptune-X: Active X-to-Maritime Generation for Universal Maritime Object Detection (NeurIPS 2025 · similarity 0.78)
- Bridging the Gap Between End-to-End and Two-Step Text Spotting (CVPR 2024 · similarity 0.78)
- Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition (CVPR 2025 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.