Efficient document image binarization using heterogeneous computing and parameter tuning

Abstract

In the context of historical document analysis, image binarization is a first important step, which separates foreground from background, despite common image degradations, such as faded ink, stains, or bleed-through. Fast binarization has great significance when analyzing vast archives of document images, since even small inefficiencies can quickly accumulate to years of wasted execution time. Therefore, efficient binarization is especially relevant to companies and government institutions, who want to analyze their large collections of document images. The main challenge with this is to speed up the execution performance without affecting the binarization performance. We modify a state-of-the-art binarization algorithm and achieve on average a 3.5 times faster execution performance by correctly mapping this algorithm to a heterogeneous platform, consisting of a CPU and a GPU. Our proposed parameter tuning algorithm additionally improves the execution time for parameter tuning by a factor of 1.7, compared to previous parameter tuning algorithms. We see that for the chosen algorithm, machine learning-based parameter tuning improves the execution performance more than heterogeneous computing, when comparing absolute execution times.

Historical Document Image Binarization: A Review

“…MoG mixture of Gaussians, CRF conditional random field, ML machine learning, SVM support vector machine, FCN fully convolutional network, LSTM long short-term memory Chattopadhyay [23] Supervised tuning SVM classifier chooses best binarization algorithm Xiong [176] Supervised tuning SVM chooses global threshold for image blocks Westphal [170] Supervised tuning Predicts parameters for Howe [50] Messaoud [87] Supervised tuning Detects 4 types of image noise to choose parameters Ntirogiannis [111] Unsuper. tuning Estimates local window sizes based on stroke width Boiangiu [15] Unsuper.…”

“…A similar approach was used in [176], but instead a global threshold is predicted for each block. Westphal et al [170] replace the unsupervised parameter tuning of Howe [50] with a supervised approach and report improved performance.…”

“…Westphal et al [170] performed an extensive study of a CPU/GPU implementation of Howe binarization [50], finding that the fastest configuration performs all steps on the GPU, achieving a 3.5× speedup compared to the reference implementation. They implemented a parallel algorithm to solve the graph-cut energy minimization step on the GPU because the original solver used in [50] is a serial algorithm.…”

Representative Image Selection for Data Efficient Word Spotting

“…Since the extraction of TP-R and G-R requires binarized input images, we binarize all word images using Howe's binarization algorithm [12]. We predict the required binarization parameters using the approach by Westphal et al [28], which has been trained on the dataset of the 2013 document image binarization contest [17]. After visual inspection of one image, we adjust one of the predicted parameters for the Alvermann Konzilsprotokolle dataset and the George Washington dataset to increase the recall.…”

Tool for automatic tuning of binarisation techniques

“…In order to estimate House's method parameters in a timely fashion, Westphal et al . [29] proposed to predict the parameters using multivariate regression based on four image features. Using these four image features, random forests were used to learn a mapping from these features to suitable parameters for a given image.…”