Source-linked AI summary
Combining Multiple Feature Extraction Techniques for Handwritten Devnagari Character Recognition
Sandhya Arora, Debotosh Bhattacharjee, Mita Nasipuri, Dipak Kumar Basu, Mahantapas Kundu
TL;DR
Handwritten Devnagari recognition remains challenging because of script complexity, modifiers, compound characters, and limited prior research. The paper combines multiple feature-based classifiers and reports 92.80% recognition on 4,900 samples using top-five choices.
Problem
Handwritten Devnagari recognition is challenging because of stroke variations, modifiers, loops, conjuncts, and compound characters, amid limited prior research.
Method
The system extracts character features, including contour-based chain-code and segmented intersection features, and combines four classifier decisions.
Results
92.80% overall recognition was obtained on 4,900 samples when considering the top five choices.
Takeaways & Limitations
The reported approach outperformed the compared recent methods and was also described as applicable to handwritten numerals and other Indian scripts.
Abstract
from arXiv · showhide
In this paper we present an OCR for Handwritten Devnagari Characters. Basic symbols are recognized by neural classifier. We have used four feature extraction techniques namely, intersection, shadow feature, chain code histogram and straight line fitting features. Shadow features are computed globally for character image while intersection features, chain code histogram features and line fitting features are computed by dividing the character image into different segments. Weighted majority voting technique is used for combining the classification decision obtained from four Multi Layer Perceptron(MLP) based classifier. On experimentation with a dataset of 4900 samples the overall recognition rate observed is 92.80% as we considered top five choices results. This method is compared with other recent methods for Handwritten Devnagari Character Recognition and it has been observed that this approach has better success rate than other methods.
I. INTRODUCTION
The paper addresses the limited research on offline handwritten Devnagari recognition, a challenging task because of script complexity and connected character structures. It proposes a feature-based recognition system for Devnagari characters.
- Devnagari recognition remains comparatively under-researched despite the script’s broad use across major Indian languages.
- The proposed system recognizes offline handwritten Devnagari characters using shadow, intersection, and chain code histogram features.
- Recognition is difficult because characters contain loops, conjuncts, modifiers, disconnected or multistroke components, and connected header lines.
II. PROPOSED METHOD
The proposed method converts handwritten character images into binary representations and extracts multiple structural and contour-based features. Its preprocessing includes iterative threshold refinement before feature extraction.
- It extracts intersection, shadow, and chain code histogram features from scaled character images.Intersection features use thinned segmented images, shadow features use octants, and chain code histograms use contour information.
- The proposed technique is illustrated as a block diagram of the recognition pipeline.
- The method first converts grayscale character images into binary foreground-background representations using a dynamically refined threshold.The threshold is iteratively updated from foreground and background means until successive changes fall below 2%.
B. Scaling of the binary character images
Character images are tightly cropped and resized to a standard 100 × 100-pixel representation. Morphological closing and dilation then smooth contours and fill some holes.
- Each character is enclosed by a tight rectangular boundary, cropped, and scaled to 100 × 100 pixels using affine transformation.
- Morphological closing and dilation smooth character contours and fill some holes after scaling.
C. Feature Extraction
The method extracts four complementary feature sets from scaled Devnagari character images, using global shadow projections and segmented structural, contour, and line-fitting representations.
- Shadow features: 16 shadow features measure character projections on two perpendicular sides of each of eight image octants.These features are computed globally on the scaled character image.
- Chain-code features: 200 chain-code histogram features count contour-direction frequencies across 5 × 5 blocks.Contour points are detected from 4-connected background neighbors, then chain-coded and represented with eight direction frequencies per block.
- Intersection features: 32 intersection features record open ends and junctions across 16 segments of a thinned, scaled character skeleton.The first 16 features represent open ends and the remaining 16 represent junction counts; thinning reduces redundant detections.
- Straight-line fitting: 48 line-fitting features describe straight lines fitted in 16 image segments using intercepts and two slope-derived features.Each segment contributes one intercept, one f1 value, and one f2 value to the representation.
III. DEVNAGARI CHARACTER RECOGNITION
Recognition uses four separate three-layer MLP classifiers, each matched to one extracted feature set. The classifiers use backpropagation with sigmoid activations and feature-dependent input sizes.
- Classifier architecture: Four three-layer MLPs classify characters using 32 intersection, 16 shadow, 48 line-fitting, or 200 chain-code features.At this stage, the characters are noncompound single symbols, so segmentation is not required.
- Training: Backpropagation trains the classifiers by minimizing squared training error with gradient descent and sigmoid activation.The learning rate is 0.8 and the momentum term is 0.7.
A. Classifier Combination
The system combines four MLP classifiers trained on different feature sets using weighted majority voting. The weights reflect the relative classifier decisions and reported individual performance.
- Weighted majority voting combines decisions from four MLP classifiers trained on different feature sets.The classifier combination is presented as a way to combine outputs from similar neural networks trained on different features.
- The combined decision assigns class support by weighting each classifier's output across the 50 classes.The reported weights are 0.349, 0.197, 0.326, and 0.128, ordered according to d1 > d3 > d2 > d4.
- 64.90% was obtained by the classifier trained with chain code histogram features.
- 36.71% was obtained by the classifier trained with intersection features.
- 60.59% was obtained by the classifier trained with shadow features.
- 24.83% was obtained by the classifier trained with straight line fitting features.
IV. RESULTS
The evaluation uses 4,900 handwritten Devnagari character samples with separate training and test sets and reports recognition across top-choice settings. The authors report 92.80% accuracy and compare it with earlier results.
- The dataset contains 4,900 samples, divided into 3,332 training samples and 1,568 test samples.The study reports top 1 through top 5 classification choices and uses 3-fold cross-validation.
- 92.80% accuracy was obtained for the authors' handwritten Devnagari character recognition method, compared with 90.65% for an earlier coarse classification result.The evaluation used 4,900 samples, exceeding the 4,750 samples used in the cited earlier study.
- A maximum-response classification strategy achieved 80.71% accuracy and was reported as better than results from two cited methods.
V. CONCLUSION
The paper proposes off-line handwritten Devnagari character recognition using three feature types and simple feedforward MLPs. It identifies possible applications to numerals and other Indian scripts.
- The proposed system combines three feature types with simple feedforward Multilayer Perceptrons for off-line handwritten Devnagari character recognition.
- The technique is identified as applicable to handwritten Devnagari numerals and potentially useful for research on other Indian scripts.