Realistic License Plate Restoration and Recognition Dataset (RLPR)

Description

[Dataset Description] This dataset was constructed to facilitate the development of methods aimed at improving image quality and recognition accuracy of license plates captured from dashcam videos. It contains paired low-quality license plate video sequences and corresponding high-quality target images, enabling quality enhancement model training and evaluation. Each sample includes a sequence of 31 low-quality license plate frames alongside a single high-quality license plate image that serves as the target (pseudo-ground truth) for quality enhancement. This dataset can be used for training models to enhance image quality and includes recognition labels for assessing recognition accuracy. Additionally, the dataset provides outputs obtained using the MF-LPR² model proposed in our paper, offering a benchmark for performance comparison. [Dataset Contents] 1) Low-Quality License Plate Image Sequences - 31 frames per sample 2) High-Quality License Plate Images (Pseudo-Ground Truth) - Includes images with and without ROI cropping - ROI coordinates provided 3) License Plate Recognition Label - For privacy issue, we only provide numerical values 4) MF-LPR² Model Super-Resolution (SR) Results - Output of the proposed MF-LPR² model 5) Homography Transformation Coordinates - Coordinates for aligning low- and high-quality images - Utility tool is also provided

Steps to reproduce

1. Collection of Low-Quality Sequences: We extracted 31-frame sequences from 1,052 dash cam videos, selecting frames with visible license plates that were sufficiently low in quality for restoration. Only sequences containing low-quality license plates that require enhancement were included. 2. Generation of High-Quality Reference Images (Pseudo GT): From each sequence, we selected the frame with the highest quality license plate as a "pseudo ground truth (GT)" image. This serves as the target for quality enhancement of the low-quality sequence. 3. License Plate Detection and Cropping: We used a DeepLabV3 model, based on a ResNet-101 backbone and fine-tuned on 130 manually annotated road images, to detect the license plate regions in each frame. We then manually refined the detected regions and expanded the bounding boxes to ensure no important details were missed during cropping. 4. Alignment and Transformation of Pseudo GT Images: To align the high-quality reference image with the low-quality frames, we applied a homography transformation by marking the four corner points of the license plates in both the pseudo GT and the low-quality frames. This allowed us to accurately align the high-quality image with the target frames, focusing only on the license plate regions during evaluation.

Institutions

Categories

Funding

Licence