Cluster-Based Sample Selection for Document Image Binarization

Abstract

The current state-of-the-art, in terms of performance, for solving document image binarization is training artificial neural networks on pre-labelled ground truth data. As such, it faces the same issues as other, more conventional, classification problems; requiring a large amount of training data. However, unlike those conventional classification problems, document image binarization involves having to either manually craft or estimate the binarized ground truth data, which can be error-prone and time-consuming. This is where sample selection, the act of selecting training samples based on some method or metric, might help. By reducing the size of the training dataset in such a way that the binarization performance is not impacted, the required time spent creating the ground truth is also reduced. This thesis proposes a clusterbased sample selection method, based on previous work, that uses image similarity metrics and the relative neighbourhood graph to reduce the underlying redundancy of the dataset. The method is implemented with different clustering methods and similarity metrics for comparison, with the best implementation being based on affinity propagation and the structural similarity index. This implementation manages to reduce the training dataset by 46% while maintaining a performance that is equal to that of the complete dataset. The performance of this method is shown to not be significantly different from randomly selecting the same number of samples. However, due to limitations in the random method, such as unpredictable performance and uncertainty in how many samples to select, the use of sample selection in document image binarization still shows great promise.

Document Analysis and Recognition: A survey

“…GiB [65]: Extracts features for clustering using game theory Shallow ML Kasmin [66]: Ensemble of 8 SVMs #: Hamza [67], Rabelo [68], Pastor [69] Deep Learning FCNs: Pastor [70], Calvo-Zaragoza [71],PDNet [72], Morphological networks with grayscale dilation and erosion operations: Mondal [73], U-Net architecture with attention layers: Kang [74]; GANs: Tensmeyer [75], Zhao [76]; LSTMs: Afzal [77], Westphal [78] LSTMs + Deep learning methods outperform a non-convolutional deep MLP (including histogram and median filter features; − Less availability of labelled data for training; Recurrent networks (LSTM, BI-LSTM) not as effective as FCN; # Bhunia [79], Krantz [80] Parameter tuning Supervised tuning: Xiong [81], Messaoud [82]; Unsupervised tuning: Ntirogiannis [83], Liang [84], + Finds best parameter settings for algorithms, − Output of algorithm dependant on these parameter settings + Pros, − Cons, # Related research layout analysis procedure, as the document types are varied. With this information, there are two levels of segmentation.…”

Historical Document Image Binarization: A Review

“…Bhunia et al [14] proposed a unique approach to train a binarization network using unpaired training data (i.e., the grayscale and binary images do not correspond) and achieved an impressive 97.8% FM on DIBCO13. To reduce the amount of needed training data, Krantz and Westphal [68] proposed a clustering method to ensure only diverse data are labeled. This reduces dataset size by 50% at a modest loss in accuracy.…”

“…There is also limited labeled training data for deep models. Though some initial explorations around this issue have been made [14,55,68,165], there is still a large gap to address.…”

Representative Image Selection for Data Efficient Word Spotting

“…However, one drawback of this approach is that it relies on label information in the original training set. Krantz and Westphal [13] follow a similar strategy for selecting training samples for image binarization using a relative neighborhood graph. In contrast to Rayar et al, their approach avoids the need for a labeled set to choose from by creating pseudo labels through clustering.…”