A Holistic Approach for Handwritten Hindi Word Recognition

Abstract

Holistic word recognition attempts to recognize the entire word image as a single pattern. In general, it performs better than segmentation based word recognition model for known, fixed and small sized lexicon. The present work deals with recognition of handwritten words in Hindi in holistic way. Features like area, aspect ratio, density, pixel ratio, longest run, centroid and projection length are extracted either from entire word image or from the hypothetically generated sub-images of the same. An 89-elements feature vector has been designed to represent each word in the feature space and five different classifiers have been used for measuring recognition performances. Considering the complexities of Hindi characters, the technique shows an impressive result using a Multilayer Perceptron (MLP) based classifier. Moreover, the technique shows scale and rotation invariant nature to a significant extent.

H‐WordNet: a holistic convolutional neural network approach for handwritten word recognition

“…In this section, we have compared the performance of the proposed H‐WordNet model with state‐of‐the‐art methods. For comparison, we have considered the holistic word recognition approaches reported in [4, 17–19, 36, 38]. Also, the efficacy of two deep learning models such as deep stacked autoencoder [53] and fire module based CNN [54] is compared.…”

“…They have performed classification using a combination of different classifiers such as multi‐layer perceptron (MLP), SVM, and extreme learning machine (ELM), and achieved satisfactory results on IFNENIT database containing Arabic handwritten city names of 21 different classes. Malakar et al [36] proposed a holistic system for handwritten word recognition in Devanagari script. They have extracted features such as area, aspect ratio, density, pixel ratio, longest run, centroid, and projection length from the word images.…”

An iterative approach to unsupervised outlier detection using ensemble method and distance-based data filtering

“…The F-score [37,38], also called the F1-score or the F-measure reflects upon a test's accuracy. It is defined as the harmonic mean of the precision and recall values.…”

Offline hand-drawn circuit component recognition using texture and shape-based features

“…So we prepare an in-house dataset containing 20 different circuit components which are listed in Table 1. We collect 150 samples for each circuit component that are drawn by different individuals who happen to be engineering students, faculty members and research scholars in a pre-formatted datasheet similar to the works [7,21]. In total, we make a dataset containing 3000 sample images of 20 different circuit components (150 per circuit component).…”

“…Let the number of foreground pixels on the contour of an image is N. We take (x(n), y(n)) as a point on the contour, where n ∈ [1, N]. The centroid distance feature [21] first calculates the geometric centroid, let (g x , g y ), of the contour image and then the distance of the point (x(n), y(n)) from centroid is r(n) which is calculated by eq. (1).…”