Benchmarks & Experimental Results
LiteALPR was rigorously benchmarked on a comprehensive multi-national dataset totaling 52,595 images:
- Brazil (RodoSol-ALPR): 20,000 images (with 9,560 evaluated in test set)
- China (CBLPRD-330k): 20,000 images (with 9,450 evaluated in test set)
- Vietnam: 12,595 images (public dataset + self-collected traffic footage, with 5,990 evaluated in test set)
For the text recognition task, the full corpus of 52,595 localized plate crops was used. To simulate challenging real-world environments and ensure the OCR engine's resilience, half (50%) of these crops were artificially distorted to emulate out-of-focus optics and high-speed motion blur. The recognition dataset was distributed into:
- Training set: 25,000 crops
- Validation set: 2,595 crops
- Test set: 25,000 crops (comprising 9,560 from Brazil, 9,450 from China, and 5,990 from Vietnam)
All benchmarks were evaluated in native ONNX FP32 precision, reporting latency and throughput as \(\text{mean} \pm \text{std}\) evaluated over 1,000 real-world test images or cropped plates across 5 consecutive execution passes.
1. End-to-End Pipeline Performance
Comparison against existing state-of-the-art and lightweight open-source ALPR pipelines:
| Platform | Model / Pipeline | Pre (ms) | Infer (ms) | Post (ms) | Total (ms) | Throughput (FPS) | Accuracy (%) |
|---|---|---|---|---|---|---|---|
| GPU (RTX 3060 12GB) | YOLOv8n + SVTRv2 (Baseline) | 1.54 ± 0.41 | 12.38 ± 2.43 | 1.55 ± 0.20 | 15.84 ± 2.79 | 63.1 ± 0.7 | 88.60% |
| fast-alpr | 4.58 ± 0.85 | 15.06 ± 3.27 | 5.62 ± 2.96 | 25.27 ± 5.67 | 40.1 ± 4.3 | 22.67% | |
| LiteALPR (Ours) | 1.63 ± 0.40 | 11.57 ± 5.87 | 1.49 ± 0.12 | 15.05 ± 6.01 | 66.5 ± 1.8 | 89.15% | |
| CPU (Ryzen 5 4600G) | YOLOv8n + SVTRv2 (Baseline) | 5.26 ± 3.55 | 58.53 ± 13.50 | 0.95 ± 0.78 | 65.08 ± 14.28 | 15.4 ± 0.3 | 88.60% |
| fast-alpr | 5.10 ± 1.18 | 24.02 ± 3.29 | 4.17 ± 2.44 | 33.30 ± 4.92 | 30.1 ± 0.8 | 22.67% | |
| LiteALPR (Ours) | 6.52 ± 3.99 | 34.04 ± 9.32 | 1.80 ± 2.06 | 42.67 ± 8.66 | 23.4 ± 0.2 | 89.15% |
Key Takeaway: While alternative lightweight pipelines such as
fast-alprdegrade down to 22.67% accuracy under blur, low light, and tilt angles, LiteALPR preserves 89.15% sequence accuracy while maintaining ultra-high throughput (~66.5 FPS on GPU, ~23.4 FPS on CPU).
2. Detection Module Ablation (YOLOv8n-Efficient)
Impact of progressively replacing C2f blocks with lightweight C3Ghost modules across head and backbone:
| Configuration | Parameters (M) | GFLOPs | GPU Latency (ms) | CPU Latency (ms) | \(\text{mAP}_{50}\) (%) | \(\text{mAP}_{50-95}\) (%) |
|---|---|---|---|---|---|---|
| Baseline YOLOv8n | 3.01 | 8.19 | 10.02 ± 1.29 | 40.92 ± 5.44 | 99.45% | 88.48% |
| Head (C3Ghost) | 2.52 | 7.19 | 9.54 ± 1.24 | 19.69 ± 4.26 | 99.44% | 89.33% |
| Backbone (C3Ghost) | 2.49 | 6.70 | 9.04 ± 0.51 | 19.59 ± 3.38 | 99.45% | 89.20% |
| YOLOv8n-Efficient (Ours) | 2.00 | 5.69 | 9.02 ± 0.98 | 17.45 ± 1.62 | 99.45% | 88.41% |
Replacing C2f with C3Ghost across both backbone and head achieves a 2.35× speedup on CPU (~\(40.92 \rightarrow\) ~17.45 ms) and 33.5% parameter reduction (\(3.01\text{M} \rightarrow 2.00\text{M}\)) with zero accuracy drop (maintains 99.45% \(\text{mAP}_{50}\)).
3. Recognition Module Ablation (SVTR26-Tiny)
Impact of replacing 2D attention matrices with Height-wise Average Pooling (HAP) and data synthesis augmentation:
| Decoder Architecture | Synthetic Degradation | Params (M) | GPU Lat. (ms) | CPU Lat. (ms) | Accuracy (%) | CER (%) |
|---|---|---|---|---|---|---|
| Baseline (Full Attention) | No | 5.11 | 5.55 ± 0.92 | 10.97 ± 1.26 | 86.71% | 4.13% |
| Baseline (Full Attention) | Yes | 5.11 | 5.85 ± 0.89 | 10.84 ± 0.76 | 88.25% | 3.56% |
| w/o Self-Attention | No | 4.33 | 5.95 ± 1.16 | 9.79 ± 0.74 | 85.98% | 4.29% |
| w/o Self-Attention | Yes | 4.33 | 5.58 ± 0.77 | 9.81 ± 0.73 | 88.52% | 3.48% |
| HAP (Ours) | No | 4.22 | 5.29 ± 1.19 | 9.05 ± 1.34 | 86.55% | 4.16% |
| HAP + Degrade (Ours) | Yes | 4.22 | 5.03 ± 0.51 | 8.32 ± 0.98 | 89.15% | 3.28% |
Height-wise Average Pooling (HAP) completely replaces 2D attention matrices, slashing CPU recognition latency from ~10.84 ms down to ~8.32 ms while maintaining superior sequence recognition accuracy of 89.15% and lowering Character Error Rate (CER) to 3.28%.
4. Cross-Regional Generalization Benchmark
Evaluation of SVTR26-Tiny against widely-adopted sequence recognition baselines across 3 diverse geographic subsets (25,000 total test samples: 5,990 from Vietnam, 9,560 from Brazil, and 9,450 from China).
Alphanumeric Vocabulary: All evaluated recognition models transcribe standard Latin uppercase alphabets and numeric digits (A–Z, Đ, 0–9, totaling 37 alphanumeric characters plus 1 CTC blank token). For multi-line plates (e.g. Vietnam and Brazil formats), Chinese provincial Hanzi characters in CBLPRD-330k were normalized or evaluated on the alphanumeric sequence, ensuring a fair, unified Latin-character benchmark across all international datasets.
The table below presents the performance of the evaluated lightweight architectures, ranked in progressive order of regional generalization:
| No. | Model | Vietnam Acc (%) | Brazil Acc (%) | China Acc (%) | Vietnam CER (%) | Brazil CER (%) | China CER (%) |
|---|---|---|---|---|---|---|---|
| 1 | CRNN | 35.17% | 55.96% | 67.88% | 36.28% | 18.31% | 11.64% |
| 2 | PP-OCRv4-Mobile | 64.50% | 87.28% | 92.05% | 13.00% | 3.02% | 1.84% |
| 3 | TrOCR | 70.20% | 90.85% | 91.45% | 11.44% | 2.50% | 2.03% |
| 4 | SVTRv2-Tiny | 81.36% | 90.54% | 96.00% | 5.05% | 2.19% | 0.82% |
| 5 | SVTR26-Tiny (Ours) | 82.93% | 90.98% | 96.51% | 4.69% | 2.06% | 0.71% |
Summary: SVTR26-Tiny attains top performance among the evaluated architectures on the challenging Vietnam benchmark (82.93%) and China benchmark (96.51% with 0.71% CER), while achieving the highest 90.98% accuracy in Brazil.