Source-linked AI summary
Diagonal Based Feature Extraction for Handwritten Alphabets Recognition System using Neural Network
J. Pradeep, E. Srinivasan, S. Himavathi
TL;DR
The paper addresses offline handwritten-character recognition, where scanned images must be transformed into features for accurate classification. It introduces diagonal feature extraction combined with a multilayer feed-forward backpropagation network. The diagonal approach achieves higher recognition accuracy than horizontal and vertical extraction methods, with the reported best result reaching 97.8%.
Problem
Offline handwriting recognition must classify completed scanned images despite lacking the temporal stroke information available in online recognition.
Method
The system preprocesses and segments scanned characters, extracts diagonal features from 90x60-pixel zoned images, and classifies them with a multilayer feed-forward backpropagation network.
Results
97.8% recognition accuracy is reported for the diagonal feature-extraction system, which outperforms the horizontal and vertical alternatives.
Takeaways & Limitations
The proposed diagonal feature-extraction system is presented as suitable for converting handwritten documents into structural text and recognizing handwritten names.
Abstract
from arXiv · showhide
An off-line handwritten alphabetical character recognition system using multilayer feed forward neural network is described in the paper. A new method, called, diagonal based feature extraction is introduced for extracting the features of the handwritten alphabets. Fifty data sets, each containing 26 alphabets written by various people, are used for training the neural network and 570 different handwritten alphabetical characters are used for testing. The proposed recognition system performs quite well yielding higher levels of recognition accuracy compared to the systems employing the conventional horizontal and vertical methods of feature extraction. This system will be suitable for converting handwritten documents into structural text form and recognizing handwritten names.
1. INTRODUCTION
Offline handwriting recognition is challenging because completed writing is available only as an image, making preprocessing, segmentation, and feature extraction central to recognition performance. The paper proposes diagonal feature extraction with a feed-forward backpropagation network to improve accuracy and training efficiency.
- Recognition context: Offline recognition uses completed scanned images, unlike online recognition, which also provides stroke order and temporal coordinates.Online methods have been reported as superior because of this temporal information, while neural networks can still achieve comparatively high accuracy offline.
- Recognition context: Preprocessing, segmentation, and feature extraction shape character images into inputs suitable for neural-network training and recognition.Segmentation isolates characters and resizes them to m x n pixels before training.
- Prior methods: Feature extraction is a major determinant of recognition performance, with established alternatives including zoning, projection histograms, geometric moments, Fourier descriptors, and gradient features.The paper positions its diagonal scheme within a broad set of previously used extraction methods.
- Prior methods: Earlier systems reported 94% accuracy for handwritten English characters using boundary tracing and Fourier descriptors, and 90.50% for handwritten Kannada numerals using horizontal/vertical strokes and endpoints.The Kannada approach used thinning, which the cited work says can cause feature loss.
- Paper approach: The proposed method extracts diagonal features from resized character zones and feeds them to a backpropagation network, with studies reporting good accuracy and less training time.The paper uses 90x60-pixel characters divided into 54 zones and extracts 54 features in the described procedure.
2. THE PROPOSED RECOGNITION SYSTEM
The proposed recognition system processes scanned handwriting through acquisition, preprocessing, segmentation, feature extraction, classification, recognition, and postprocessing. Its preprocessing binarizes, detects edges, dilates, and fills holes before isolated characters are resized for recognition.
- System pipeline: The recognition pipeline comprises preprocessing, segmentation, feature extraction, classification and recognition, and postprocessing stages.The paper presents this pipeline schematically as the proposed recognition system.
- System pipeline: Image acquisition accepts a scanned input from a scanner, digital camera, or other digital device in formats such as JPEG or BMT.The acquired image is the input to subsequent processing stages.
- Preprocessing: Preprocessing converts grayscale images to binary images, detects edges with Sobel filtering, dilates the image, and fills holes.These operations enhance the scanned image so it is suitable for segmentation.
- Segmentation: Segmentation decomposes character sequences into isolated characters by labelling, then uniformly resizes each character to 90X60 pixels.Labelling also provides the number of characters in the input image.
3. PROPOSED FEATURE EXTRACTION METHOD
The paper extracts diagonal features from zoned character images by aggregating foreground pixels along diagonal lines. This produces compact feature values for recognition while also allowing additional rowwise and columnwise features.
- Diagonal extraction: Each 90x60-pixel character is divided into 54 equal 10x10-pixel zones for diagonal feature extraction.The procedure is applied sequentially to every zone.
- Diagonal extraction: Each zone contains 19 diagonal lines, whose foreground-pixel sums are averaged to form one feature value.Empty diagonals contribute zero-valued features for the corresponding zones.
- Feature representations: The diagonal procedure yields 54 features per character, with additional rowwise and columnwise averaging producing 69-feature representations.The passage states that 9 and 6 additional features are obtained by averaging values.
4. CLASSIFICATION AND RECOGNITION
The classification system uses extracted feature vectors as inputs to a multilayer feed-forward backpropagation network. Two hidden layers with 100 neurons each and a competitive output layer perform character classification and recognition.
- Network architecture: A feed-forward backpropagation network with architecture 54-100-100-38 performs classification using the extracted feature vector.The input size corresponds to the number of features, while the output size corresponds to the total number of character classes.
- Network architecture: The two hidden layers use log sigmoid activation, and the competitive output layer identifies one character.The most compact hidden-layer configuration was selected by trial and error.
- Training setup: The network training setup supports 54/69 input nodes, 100 hidden nodes per layer, and 38 output nodes for alphabets, numerals, and special symbols.Training uses gradient descent with momentum, adaptive learning, mean square error, and a training goal of 0.000001.
- Training setup: The network architecture is illustrated as a three-layer neural network for character recognition.The figure is associated with the paper’s classification and recognition stage.
5. RESULTS AND DISCUSSION
The system evaluates three feature-extraction orientations with 54- and 69-feature representations, emphasizing recognition accuracy, convergence speed, and training stability. Diagonal extraction provides the strongest reported recognition performance, while the 69-feature configuration reaches the training goal in 854 epochs.
- Evaluation: The trained recognition system is tested on unknown datasets after networks with different feature-vector lengths are built.The number of input nodes is selected according to the feature count.
- 54 features: The 54-feature comparison evaluates recognition accuracy, convergence speed, and training stability across the three extraction orientations.For diagonal extraction, the desired performance goal is achieved in 923 epochs.
- 69 features: 69 features including rowwise and columnwise features yield higher recognition accuracy with diagonal extraction than the other orientations.The 69-feature diagonal network reduces mean square error to the desired level in 854 epochs.
6. IMPLEMENTATION ON GRAPHICAL USER INTERFACE
A menu-based graphical user interface supports the recognition workflow and lets users choose feature-extraction methods. It also provides controls for preprocessing, training, test-image selection, recognition, and displaying recognized output.
- Interface design: The interface provides two menus: one for five processing stages and another for selecting the feature-extraction type.The feature-extraction options are presented through the menu-based GUI.
- Recognition workflow: Users can preprocess images, select and apply a feature-extraction method, and train the neural network through the GUI.These functions are exposed as menu-based operations rather than command-line steps.
- Recognition workflow: After training, users select test images, initiate recognition, view recognized images in a notepad, and repeat the process for additional images.An exit control closes the system after test-image recognition is complete.
7. CONCLUSION
The paper concludes that diagonal feature extraction improves off-line handwritten English alphabet recognition, with the 69-feature configuration producing the highest reported accuracy. The system is presented as suitable for several document and address-recognition applications.
- Recognition performance: Diagonal feature extraction achieves the highest reported recognition accuracy: 97.8% with 54 features and 98.5% with 69 features.The comparison includes horizontal and vertical feature-extraction methods.
- Recognition performance: The 69-feature representation performs better than the 54-feature representation across the evaluated feature-extraction types.The 69-feature approach includes rowwise and columnwise features.
- Applications: The proposed off-line system is intended for postal or parcel address recognition, bank processing, document reading, and handwritten-document conversion.The conclusion describes these as suitable applications for the system.