Deep Restoration of Vintage Photographs From Scanned Halftone Prints

Qifan Gao, Xiao Shu, Xiaolin WuICCV 2019 · OpenReview

Abstract

A great number of invaluable historical photographs unfortunately only exist in the form of halftone prints in old publications such as newspapers or books. Their original continuous-tone films have long been lost or irreparably damaged. There have been attempts to digitally restore these vintage halftone prints to the original film quality or higher. However, even using powerful deep convolutional neural networks, it is still difficult to obtain satisfactory results. The main challenge is that the degradation process is complex and compounded while little to no real data is available for properly training a data-driven method. In this research, we adopt a novel strategy of two-stage deep learning, in which the restoration task is divided into two stages: the removal of printing artifacts and the inverse of halftoning. The advantage of our technique is that only the simple first stage requires unsupervised training in order to make the combined network generalize on real halftone prints, while the more complex second stage of inverse halftoning can be easily trained with synthetic data. Extensive experimental results demonstrate the efficacy of the proposed technique for real halftone prints; the new technique significantly outperforms the existing ones in visual quality.

Introduction

Before the digital age, mass-producing photographs in newspapers or books required halftoning—a process that uses discrete black ink dots of varying sizes to simulate continuous shades of gray. For countless historical records, the original film negatives are lost, leaving these grainy, ink-dotted prints as the only surviving visual evidence.

Restoring these images, a process known as "inverse halftoning," is notoriously difficult for deep learning models. Real vintage scans suffer from complex degradations like ink smears, paper fiber textures, and physical wear. Furthermore, there is a severe lack of "paired data" (original photos vs. degraded scans) required for standard supervised training.

The Two-Stage Restoration Framework

To bridge the gap between messy physical prints and clean digital restoration, this paper proposes a decoupled two-stage pipeline.

Stage 1: Artifacts Removal Subnet ($N_s$)

The first stage acts as a "cleaner." It maps the scanned, degraded halftone image to a "clean" synthetic-looking halftone. Because paired data is unavailable, this stage utilizes unsupervised learning with an adversarial loss (similar to CycleGAN). The goal is to make real scans statistically indistinguishable from clean, digitally generated halftones while preserving the underlying image content and dot structure.

Stage 2: Inverse Halftoning Subnet ($N_i$)

Once the image is converted into a clean digital halftone format, the second stage performs the actual restoration into a continuous-tone image. This subnet is trained using massive amounts of synthetic paired data (original images paired with their digital halftone versions). By using Residual Residual Dense Blocks (RRDB), the network effectively recovers high-frequency details and natural textures.

The two-stage restoration process showing (a) the original scanned halftone, (b) the cleaned intermediate halftone, and (c) the final restored continuous-tone image.

Experimental Results

The proposed method was evaluated on a real-world test set of 379 vintage photo prints.

Qualitative Performance

Visual results on 1920s-era scans show a dramatic improvement over existing methods. While traditional inverse halftoning often leaves "ghost" dot patterns or produces blurry results, this two-stage approach yields sharp edges and natural-looking skin textures.

Detailed view of the final restored image, highlighting the recovery of natural textures in the eye and hair regions.

Quantitative and Downstream Impact

On synthetic benchmarks, the model achieved a PSNR of 30.10, surpassing state-of-the-art models like ESRGAN. To validate the real-world utility, the researchers tested the restoration on face detection tasks. Using the MTCNN algorithm, face detection accuracy improved from 60.1% on raw scans to 75.7% on restored images, significantly outperforming other restoration techniques.

Conclusion

By decoupling the restoration task into unsupervised cleaning and supervised inverse halftoning, this framework successfully overcomes the domain gap inherent in vintage photograph restoration. While challenges remain regarding severe physical damage and color restoration, this approach provides a powerful tool for preserving historical visual archives.

Summary updated

Abstract

A great number of invaluable historical photographs unfortunately only exist in the form of halftone prints in old publications such as newspapers or books. Their original continuous-tone films have long been lost or irreparably damaged. There have been attempts to digitally restore these vintage halftone prints to the original film quality or higher. However, even using powerful deep convolutional neural networks, it is still difficult to obtain satisfactory results. The main challenge is that the degradation process is complex and compounded while little to no real data is available for properly training a data-driven method. In this research, we adopt a novel strategy of two-stage deep learning, in which the restoration task is divided into two stages: the removal of printing artifacts and the inverse of halftoning. The advantage of our technique is that only the simple first stage requires unsupervised training in order to make the combined network generalize on real halftone prints, while the more complex second stage of inverse halftoning can be easily trained with synthetic data. Extensive experimental results demonstrate the efficacy of the proposed technique for real halftone prints; the new technique significantly outperforms the existing ones in visual quality.
BibTeX
@inproceedings{gao2019deep,
  author = {Qifan Gao and Xiao Shu and Xiaolin Wu},
  title = {{Deep Restoration of Vintage Photographs From Scanned Halftone Prints}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2019},
  url = {https://openreview.net/forum?id=ZggnI9A6oH},
}
Figure