Restoration of Old Digitized Autochrome Images using Deep Learning Techniques

Autochromes are an early method of color photography from the beginning of the 20th century. The technical characteristics of Autochrome plates make them very vulnerable to damage. One of these damages is the so-called greening, in which green color bleeds inside the Autochrome plates due to moisture. In this thesis, methods for deep learning based digital restoration of these defects are tested and compared with the results of a manual digital restoration using Adobe Photoshop. For this purpose, two synthetic datasets are created to serve as a training basis. These datasets will later be used to train different variations of the Pix2Pix model, the CycleGAN model, and the Channel Interaction Restoration (ChaIR) model. The ChaIR model architecture can already achieve state-of-the-art (SOTA) results for similar tasks, such as image dehazing. The model training results initially show that the first, simpler dataset generated based on the image composition is insufficient. The second dataset, which simulates the nature of the defects in greater depth and through direct color manipulation, provides results that can be transferred to real greening defects. The objective assessment of the restored Autochromes using no-reference metrics such as the Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) and the Naturalness Image Quality Evaluator (NIQE) score also appears insufficient. After several different evaluations using synthetic and real image data, one approach stands out. In this approach, a ChaIR model pretrained for the dehazing of outdoor images was finetuned with the synthetic image data using transfer learning. The resulting model can identify greening defects on Autochromes and adjust their colors. The resulting defects appear optically less recognizable but are still identifiable and appear slightly discolored. The underlying structures of the image are preserved after being processed by the model. The deep learning based methods offer various advantages and disadvantages compared to restoration methods with Photoshop. All in all, both methods together have the potential to ensure the best possible restoration process for Autochromes that are affected by greening.