Restoration of multispectral images of ancient documents

Abstract

This thesis is concerned with the restoration of images of historical documents. The ancient writings imaged contain partially faded-out characters or are degraded by background variations. MultiSpectral Imaging (MSI) has proven to be a valuable tool for the non-invasive investigation of such ancient manuscripts, since it can be used to acquire information that is invisible to the human eye. The document images which are examined in this work, have been acquired with a portable MSI system. The imaging in narrowband spectral ranges led to a considerable legibility increase. The images taken form the basis for two kinds of restoration techniques that are introduced in this work: First, an enhancement method is proposed that projects the multispectral samples on a lower dimensional space by applying an Linear Discriminant Analysis (LDA) based transformation. Thus, not only the dimensionality of the multispectral images is lowered, but also the legibility of the degraded writings is increased. A qualitative analysis conducted by philologists shows that the method partially outperforms unsupervised dimension reduction methods, which are used in previous works. The second aim of this work is the separation of the ancient writings from the remaining background. Such binarization methods are used as a preprocessing step for other document image analysis methods, including OCR (OCR) or writer identification. Multiple binarization methods have been developed for the multispectral document images considered: Two methods make use of a target detection algorithm, which is used to determine if ink is present within the multispectral samples. A further binarization method is introduced, which makes use of Gaussian Mixture Model (GMM) based clustering. The methods introduced make use of spatial and spectral information. Furthermore, a Fully Convolutional Network (FCN) is used for the binarization task. The methods are evaluated on two databases: First, the methods are applied on the MultiSpectral Text Extraction (MS-TEx) dataset, where the methods achieve promising results. The best performances are gained by the target detection-based methods. These methods participated in the MS-TEx 2015 contest, where they were ranked first and second. Second, the methods are evaluated on the MultiSpectral Document Binarization (MSBin) dataset. This dataset is larger and allows for a successful training of the FCN, which outperforms the remaining binarization methods. Nevertheless, the results gained by all methods proposed are superior to the results which are gained by a traditional binarization approach that is designed for grayscale images.