Source-linked AI summary
Deep Learning Autoencoder Approach for Handwritten Arabic Digits Recognition
Mohamed Loey, Ahmed El-Sawy, Hazem EL-Bakry
TL;DR
Arabic handwritten-digit recognition must accommodate substantial handwriting variation and large image databases. The paper uses stacked autoencoders to learn compressed features and classify digits on MADBase, reporting 98.5% testing accuracy and a 1.5% testing misclassification error rate.
Problem
Arabic handwritten-digit recognition faces substantial handwriting variation, while deep learning requires large image datasets for effective decisions.
Method
The paper uses a stacked autoencoder with two hidden feature layers and a softmax output classifier for Arabic handwritten-digit classification.
Results
98.5% classification accuracy was obtained on MADBase testing data.
Takeaways & Limitations
The stacked autoencoder learned digit-stroke features and achieved a 1.5% testing misclassification error rate on the MADBase benchmark.
Abstract
from arXiv · showhide
This paper presents a new unsupervised learning approach with stacked autoencoder (SAE) for Arabic handwritten digits categorization. Recently, Arabic handwritten digits recognition has been an important area due to its applications in several fields. This work is focusing on the recognition part of handwritten Arabic digits recognition that face several challenges, including the unlimited variation in human handwriting and the large public databases. Arabic digits contains ten numbers that were descended from the Indian digits system. Stacked autoencoder (SAE) tested and trained the MADBase database (Arabic handwritten digits images) that contain 10000 testing images and 60000 training images. We show that the use of SAE leads to significant improvements across different machine-learning classification algorithms. SAE is giving an average accuracy of 98.5%.
1. Introduction
Handwritten digit recognition supports practical applications, while Arabic digit recognition remains challenging because handwriting varies widely. The paper proposes stacked autoencoders to learn compressed feature representations for recognition.
- Handwritten digit recognition supports office automation, check verification, postal reading, and data entry.
- Deep learning uses hierarchical multilayer networks to extract features from input data.
- Autoencoders learn compressed hidden representations whose features improve on raw input representations.
- The paper proposes a stacked autoencoder recognition system for Arabic handwritten digits.
2. Related Work
Prior Arabic digit-recognition studies used diverse classifiers and feature designs, often reporting strong accuracy. However, their datasets and evaluation settings varied substantially, including several small databases.
- Arabic digit recognition studies used Gabor features with SVMs, Loci characteristics, neural networks, and hybrid classifiers.
- 99.85% average recognition rate was reported with Gabor features and SVMs using three scales and five orientations.
- 99% recognition rate was reported for Loci-based recognition on a database containing 600 Arabic digits.
- More than 96% recognition accuracy was reported for a back-propagation neural network on a small handwritten database.
- 88% overall testing accuracy was reported for a three-level classifier combining SVM, Fuzzy C Means, and Unique Pixels.
- Typewritten Arabic-digit methods reported recognition rates over 95% on some fonts.
3. Proposed Approach
The proposed approach uses stacked sparse autoencoders to learn increasingly abstract digit features, then classifies those features with a softmax layer. The architecture combines two feature-learning stages into a three-layer stacked autoencoder for Arabic handwritten digit classification.
- Motivation: The approach addresses handwriting variation and the need to learn from large Arabic handwritten-digit databases.The paper motivates an advanced recognition solution because prior methods used small databases, while deep learning requires many images.
- Autoencoder design: A stacked autoencoder consists of multiple sparse autoencoders whose outputs feed successive layers.Each autoencoder has input, encoding, and decoding layers, with the hidden layer learning an input representation.
- Autoencoder design: The autoencoder reconstructs inputs through encoder and decoder units, using learned weights, biases, and sigmoid nonlinearities.The output layer has the same number of units as the input layer.
- Stacked feature learning: The first sparse autoencoder learns primary features from raw inputs, and the second learns secondary features from those primary features.The primary representation becomes the next autoencoder’s input.
- Classification: A softmax classifier maps secondary features to digit labels.The classifier is connected after the two sparse autoencoders in the proposed pipeline.
- Final architecture: The final architecture combines two hidden feature layers with a softmax output layer capable of classifying Arabic handwritten digits.The two hidden layers represent primary and secondary features.
4. Experiment
The experiment evaluates a stacked autoencoder for Arabic handwritten digit recognition using MADBase images, with layered feature extraction followed by softmax classification. The approach achieved 98.5% testing accuracy and a 1.5% testing miss-classification error rate.
- Architecture: The stacked autoencoder uses 784-pixel inputs, hidden layers of 392 and 196 features, a softmax layer, and 10 output labels.The images contain 28×28 pixels, and the architecture is implemented in MATLAB using its deep learning toolbox.
- Dataset: The proposed system is trained and evaluated on the MADBase database, which contains 60,000 training images and 10,000 testing images.The database is used to assess recognition across varied Arabic handwritten digit images.
- Learned features: The encoder weights learn features representing digit strokes from MADBase images.Each encoder neuron has a weight vector used for feature extraction.
- Results: The SAE produced minimum misclassification error compared with a method using the same database and performed better than most cited approaches.The authors caution that comparisons with several earlier studies are difficult because those studies used smaller databases.
- Results: 98.5% classification accuracy was obtained on the testing data.The paper also reports a 1.5% testing miss-classification error rate on the MADBase benchmark.