SIDL: A Real-World Dataset for Restoring Smartphone Images with Dirty Lenses

Sooyoung Choi, Sungyong Park, Heewon KimAAAI 2025 · DOI

TL;DR

Smartphone photos often suffer from image quality degradation due to dirty lenses, a problem largely overlooked by current image restoration research. This paper introduces SIDL, a novel, real-world dataset of smartphone images with various lens contaminants, and demonstrates that existing state-of-the-art restoration models struggle to effectively handle these realistic degradations, highlighting the need for more robust techniques.

What Problem Does This Paper Address?

In our daily lives, smartphones are our primary cameras. However, the image quality captured by these ubiquitous devices can be significantly compromised by a common, yet often ignored, issue: dirty lenses. Water droplets, fingerprints, dust, and scratches can all introduce artifacts like blur, haze, and distortion, leading to degraded images. While the field of image restoration has made great strides in addressing issues like denoising, deblurring, and super-resolution, most research assumes clean or ideal lens conditions. This assumption fails to reflect the reality of everyday smartphone photography, where lens contamination is a frequent occurrence. The paper aims to bridge this gap by providing a dataset and evaluation framework specifically for restoring images affected by these realistic lens contaminants.

Methods

The core contribution of this paper is the creation of the SIDL (Smartphone Images with Dirty Lenses) dataset. To build this dataset, the authors developed a custom setup using an iPhone 12 Pro and a 3D-printed "film holder." This holder allowed them to attach thin PVC films coated with various simulated contaminants (water drops, fingerprints, dust, scratches, and mixtures) directly in front of the camera lens. They meticulously captured images under diverse lighting conditions and environments, ensuring each degraded image had a corresponding "clean" reference image. This setup allowed for the creation of realistic degradations without relying solely on synthetic generation, which can sometimes lack the nuances of real-world artifacts.

Datasets

The primary contribution is the SIDL dataset.

  • SIDL Dataset: Contains 1,588 degraded images from 300 real-world scenes, paired with clean reference images. It features five types of contaminants: dust, fingerprint, scratch, water, and mixed types. The dataset is further categorized into "Easy," "Medium," and "Hard" difficulty levels based on the severity of degradation (measured by PSNR).
  • Existing Datasets (for context): The paper briefly reviews existing datasets for image restoration tasks like deblurring (e.g., GOPRO, HIDE, RealBlur), denoising (e.g., SIDD, DND), and environmental conditions (e.g., O-HAZE, Dense-Haze). It highlights that prior "dirty lens" datasets were either limited to a single degradation type (e.g., scratch) or relied on synthetic data that may not fully capture real-world complexity.

Key Insights and Ideas

  • Real-world lens contamination is a significant challenge: The paper convincingly demonstrates that even advanced image restoration models struggle with the realistic degradations introduced by dirty lenses, suggesting this is a distinct and difficult problem.
  • Simulated contaminants can mimic real-world effects: The use of physical materials on PVC films effectively replicates the visual characteristics of lens contaminants, providing a practical approach to dataset creation.
  • Diversity of degradation is key: The SIDL dataset's strength lies in its inclusion of multiple contaminant types and their combinations, which makes it a more challenging and representative benchmark than previous single-degradation datasets.
  • Existing models lack generalization for this specific problem: The evaluation shows that models trained on general image restoration tasks or even other specific degradation datasets do not perform optimally on SIDL without fine-tuning, underscoring the unique nature of lens contamination artifacts.

Key Results

  • State-of-the-art models show limited success: When evaluated on SIDL, models like AirNet, NAFNet, Restormer, FFTformer, DiffUIR, and MambaIR achieved some level of restoration but "struggled to handle the diverse and complex contaminants in SIDL effectively."
  • DiffUIR shows promise but is not perfect: Among the tested models, DiffUIR (a diffusion-based method) consistently outperformed others across various degradation types and levels. However, even its qualitative results were described as "unsatisfactory," indicating room for improvement.
  • Dataset size is sufficient: Experiments showed that performance plateaus after training with approximately 100 scenes, suggesting the current dataset size is adequate for evaluating restoration models.
  • Pretrained models underperform: Models pretrained on other datasets performed significantly worse on SIDL compared to models fine-tuned on SIDL, reinforcing the dataset's unique challenge.

Limitations and Future Work

  • Limitations: The authors acknowledge that their dataset focuses on common contaminant types and that real-world scenarios might involve a wider, more complex range of imperfections. The simulations, while realistic, may not fully capture every nuance of actual lens degradation.
  • Future Work: The paper suggests expanding the dataset to include more varied degradation types. It also implicitly calls for the development of new, more robust, and adaptable image restoration techniques specifically designed to handle the complex and diverse artifacts introduced by dirty smartphone lenses.

Summary updated

BibTeX
@inproceedings{choi2025sidl,
  author = {Sooyoung Choi and Sungyong Park and Heewon Kim},
  title = {{SIDL: A Real-World Dataset for Restoring Smartphone Images with Dirty Lenses}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i3.32257},
}
Figure