Distinction between handwritten and machine-printed text based on the bag of visual words model

Handwritten and Machine-Printed Text Discrimination Using a Template Matching Approach

“…In terms of classification, the algorithm outperforms multi-class SVM and Random Forest approaches and in terms of features separability for each class, the approach is more robust for handling text lines and noise, and less vulnerable to the segmentation failures than the baseline Gabor features. Nevertheless, as will be shown in the experiments section, our approach outperforms the method proposed in [3] by over 13% in classification accuracy and gives similar results as in [4] despite using no labelled training data.…”

“…Because of the high similarity with the machine-printed sample, the geometric features, such as area and rectangularity used in [6], [7], [10], [12], fail to create separable classes and the samples will wrongly be classified as machine-printed. Table 1 shows the comparison of our approach with the results provided by Zagoris et al [3], [4], in which 15% of the samples in the PRImA-NHM dataset are utilised for train and the remaining 85% for the test phase. The bounding boxes-based PRImA Layout Evaluation Framework [4], [22] is used to compute overlapping regions of the classified segments with the ground truth.…”

“…Table 1 shows the comparison of our approach with the results provided by Zagoris et al [3], [4], in which 15% of the samples in the PRImA-NHM dataset are utilised for train and the remaining 85% for the test phase. The bounding boxes-based PRImA Layout Evaluation Framework [4], [22] is used to compute overlapping regions of the classified segments with the ground truth. Although the proposed PNCC algorithms do not use any training samples, their performance is only slightly lower than the one reported in [4] and ≈ 14% higher than Gabor features [3].…”

“…A novel unconventional, yet very successful approach is proposed in [3], [4]. After performing preprocessing and document segmentation, scale-invariant feature transform (SIFT) [18] is applied to each text block.…”

“…Quantitative evaluations have been conducted over the publicly available Pattern Recognition & Image Analysis Research LabNatural History Museum (PRImA-NHM) dataset [3], giving compelling results as compared with the state-of-the-art. The main contributions of the proposed HMC algorithm are as follows: 1) As opposed to [4], [5], [6], [7], [8], [9], [10], the proposed algorithm is training-free as does not rely on a trainable classifier. This provides the capability to update the gallery samples, without re-designing or modifying the decision making step.…”