This study proposes an integrated processing framework designed to enhance the readability and reliability of degraded license plate images acquired under adverse conditions such as low resolution, low illumination, motion blur, and overexposure. In r...
This study proposes an integrated processing framework designed to enhance the readability and reliability of degraded license plate images acquired under adverse conditions such as low resolution, low illumination, motion blur, and overexposure. In real-world investigative environments, CCTV and dashboard camera footage often capture license plates at small scales or with severe distortion caused by reflections, weather conditions, or viewing angles. These factors frequently lead to delays and misidentification during vehicle tracking and forensic
analysis. While previous super-resolution-based restoration methods have improved visual clarity, they inherentlygenerate non-existent pixel information, which can distort character shapes and compromise the authenticity andevidential reliability of images used in legal proceedings.
To address these challenges, this study introduces a two-stage restoration pipeline that prioritizes information preservation and combines it with a license plate–specific sequence recognition architecture. In the first stage, global contrast and structural contours weakened by noise and illumination imbalance are restored to improve plate region visibility. The YOLOv8-based detection module is then applied to automatically localize license plate areas. In the second stage, fine-grained restoration is performed exclusively on the detected regions to refine character strokes and edges. Finally, a CNN–attention-based sequence recognition model is applied to the restored license plate images, enabling robust handling of both seven- and eight-character plate formats. This sequential process is designed to selectively enhance visually relevant details while maintaining the structural consistency and evidential integrity of the original image.
To evaluate the proposed approach, a dataset was constructed using AI-Hub vehicle license plate images withsimulated degradations, including low-light, low-resolution, overexposure, blur, and composite conditions. Experimental results demonstrate that the inclusion of the restoration stages significantly improved PSNR and SSIMmetrics, as well as overall plate-level recognition accuracy compared to non-restored inputs. Through this framework, the study presents a practical and effective method to improve license plate recognition performance using existing imaging devices and data, thereby establishing a technical foundation for stable andreliable plate interpretation in traffic enforcement, criminal investigations, and digital forensic applications.