Scale Space Co-Occurrence HOG Features for Word Spotting in Handwritten Document Images

Abstract

In this paper, the authors proposed a Scale Space Co-occurrence Histograms of Oriented Gradients method (SS Co-HOG) for retrieving words from digitized handwritten documents. The poor performance of HOG based word spotting in handwritten documents is due to that HOG ignores spatial information of neighboring pixels whereas Co-HOG captures the spatial information of neighboring pixels through counting the occurrence of the gradient orientations of two or more neighboring pixels. The authors employed three scale parameter representation of an image and at each scale, they divide the word image into blocks and Co-HOG features are extracted from each block and finally concatenate them into form a feature descriptor. The proposed method is evaluated using precision and recall metrics through experimentation conducted on popular datasets such as IAM and GW and confirmed that their method outperforms for both the datasets.

HWR Text Segmentation Based on Subband Analysis

“…Для изображений текста таким структурированием является сегментация [Hochulia et al, 2018], то есть процесс разделения цифрового изображения на несколько частей. Первым этапом является сегментация текста на строки [Quang Nhat Vo et al, 2018], вторым -сегментация строк на слова [Thontadari, Prabhakar, 2016] и третьим -сегментация слов на символы [Munish et al, 2017;Benafia et al, 2017].…”

Bag of Visual Words for Word Spotting in Handwritten Documents Based on Curvature Features

“…The fourth method used for comparison is unsupervised word spotting [45] using a grid of HOG descriptors by sliding window framework. We compared the performance of proposed method with our previous work for word spotting based on Co-HOG feature descriptor extracted in scale space representation [46]. The results of existing methods are extracted from their papers where results are reported as shown in Table 3.…”

“…SIFT features are densely extracted at different patch size and then aggregated into Fisher vector representation of the word image and second method is Co-HOG feature descriptor based word spotting proposed in our previous work [46]. The results of existing methods are extracted from their papers where results are reported as shown in Table 5.…”

“…Among existing methods, [9] and [14] are based on BoVW concept and [22], [45] and [46] are non-BoVW methods. Hence we extended these non-BoVW methods to BoVW framework and evaluation results are obtained.…”