Research map / Image generation / Image restoration

LP-Diff: Towards Improved Restoration of Real-World Degraded License Plate

CVPR 2025 · Image generation › Image restoration · approach: Real-world restoration with unaligned data · 5 citations (OpenAlex)

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.

Introduces MDLP, a real-world paired multi-frame degraded license plate restoration dataset (10,245 pairs), and LP-Diff, a diffusion-based residual restoration model with inter-frame cross attention, texture enhancement, and dual-pathway fusion modules.

Problem

License plate image restoration (LPIR) methods are typically trained on synthetic degradations and fail on real-world severely degraded, low-resolution LP images captured by traffic cameras.

Approach

Collect a real-world paired multi-frame LP dataset with high-speed cameras and build a diffusion model that fuses temporal information across consecutive frames, enhances texture via Sobel-based curvature and gradient magnitude, and fuses channel and spatial attention, predicting a residual over a pretrained encoder-decoder.

Key results

• LP-Diff reaches 14.396 PSNR and 0.393 SSIM on MDLP, the best among compared methods (Real-ESRGAN 13.927 PSNR).
• On LP text recognition with a CRNN, LP-Diff achieves NED 0.198 and ACC 0.305, the only model with NED below 0.2.
• Ablation removing the Texture Enhancement Module raises NED from 0.198 to 0.289 in the ablation table (exp4 excludes DFM; TEM ablation reported as 0.264).

Measured effects reported in the paper

Surprising findings

• Pure restoration metrics (PSNR/SSIM) favor Real-ESRGAN-level or HAT-level methods in some cases, while recognition accuracy (ACC) favors diffusion methods such as ResDiff and ResShift over Real-ESRGAN for LP text.

Limitations

• Diffusion inference is slow, acknowledged by the authors.
• Dataset covers only certain regions and countries (future work to broaden).

Experimental setup

Datasets: MDLP CCPD

Benchmarks: MDLP CCPD

Metrics: PSNR SSIM FID LPIPS NED ACC

Baselines: SRCNN HAT Real-ESRGAN ResDiff ResShift

Builds on: ResDiff ResShift DDPM CRNN

Closest papers by meaning

• Towards Ship License Plate Recognition in the Wild: A Large Benchmark and Strong Baseline (AAAI 2025 · similarity 0.83)
• AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation (ICLR 2025 · similarity 0.82)
• Learning Where and What to Restore for Composite Image Restoration (NeurIPS 2026 · similarity 0.81)
• SIR-DIFF: Sparse Image Sets Restoration with Multi-View Diffusion Model (CVPR 2025 · similarity 0.81)
• Controlling Vision-Language Models for Multi-Task Image Restoration (ICLR 2024 · similarity 0.81)
• MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.81)