K. V. Prema

Generalization capability of artificial neural network incorporated with pruning method

In any real world application, the performance of Artificial Neural Networks (ANN) is mostly depends upon its generalization capability. Generalization of the ANN is ability to handle unseen data. The generalization capability of the network is mostly determined by system complexity and training of the network. Poor generalization is observed when the network is over-trained or system complexity (or degree of freedom) is relatively more than the training data. A smaller network which can fit the data will have the k good generalization ability. Network parameter pruning is one of the promising methods to reduce the degree of freedom of a network and hence improve its generalization. In recent years various pruning methods have been developed and found effective in real world applications. Next, it is important to estimate the improvement in generalization and rate of improvement as pruning being incorporated in the network. A method is developed in this research to evaluate generalization capability and rate of convergence towards the generalization. Using the proposed method, experiments have been conducted to evaluate Multi-Layer Perceptron neural network with pruning being incorporated for handwritten numeral recognition.

A Gabor filters based method for segmenting inflected characters of Kannada script

OCR system plays an important role in automatic identification of a script in a given document image, which provides important applications. A country like India, most of the people use more than one language in their day to day life; the requirement of OCR system is very much essential. There is not much work in developing OCR system for south Indian languages such as Kannada are reported in the literature. Recognition of the Kannada character is more complex and challenging, because it has large set of character with more similarity in properties among characters and characters belonging to same class have higher variability among different set of fonts. Moreover, the Kannada characters are formed by combination of basic symbols; a natural approach for recognition is to segment characters into basic symbol and recognize each symbol subsequently. Therefore a character level segmentation method is highly desirable. Also precise segmentation will certainly reduce the number of classes to recognize. The naïve use of characters images for segmentation may not yield more accurate results. With previous studies which have confirmed that the multi-channel Gabor decomposition represents an excellent tool for image segmentation and texture analysis, we propose a novel character segmentation method using Gabor filters. On comparing with manually segmented benchmark data we obtained overall accuracy of 93.82%.

Rotation invariant object recognition using Gabor filters

In recent years, Gabor filters have found effective for feature extraction as they possess many properties such as tunable to specific orientation, spectrally localized, spatially localized etc. In this paper, a rotation invariant object recognition system is proposed using Gabor filters. A set of Gabor filters are considered and directional features are extracted from an image. A Gabor Vector Set is created from an unknown image sample, which may be rotated. A combined classification approach using K-Nearest Neighbor classifier and Minimum distance classifier is developed to predict the class label of the unknown sample. Experiments are conducted on electric component images which are rotated between 0° to 360° angle. An overall recognition rate of 96.02% is observed on database of size 3971 images.

Document Image Segmentation for Kannada Script Using Zone Based Projection Profiles

In Kannada document image analysis, the document segmentation into individual lines, words and characters poses many challenges. The aksharas of Kannada language are formed by combination of glyphs of consonant, consonant conjuncts and vowel modifiers; therefore a lot variation in widths and heights of aksharas can be observed. Moreover an akshara is composed of two or more disconnected components. The consonant conjuncts pose more problems as they occasionally overlap with next line or next character. In this research we have proposed a Kannada document segmentation method using zone based projection profiles. The experiments are conducted on several scanned Kannada document with different fonts and character segmentation rate of 99.2% and consonant conjunct segmentation rate of 92.73% on an average is observed.

Multilayer Feed-Forward Artificial Neural Network Integrated with Sensitivity Based Connection Pruning Method

The Artificial Neural Network (ANN) with small size may not solve the problem while the network with large size will suffer from poor generalization. The pruning methods are approaches for finding appropriate size of the network by eliminating few parameters from the network. The sensitivity based pruning will determine sensitivity of the network error for removal of a parameter and eliminate parameters with least sensitivity. In this research a sensitivity based pruning method is integrated with multilayer feed-forward ANN and applied on MNIST handwritten numeral recognition. An analysis of effect of pruning on the network is compared with performance of a network without pruning. It is observed that the network integrated with pruning method show better generalization ability than a network without pruning method being incorporated.

Segmentation of Inflected Top Portions of Kannada Characters Using Gabor Filters

Kannada script has large number of characters with similar looking shapes among characters and characters belonging to same class have higher variability across different set of fonts. Moreover, the Kannada characters are formed by combination of basic symbols, recognition of the Kannada character is complex and challenging task. The better approach for recognition is to segment characters into basic symbol and recognize each symbol subsequently. Therefore a character level segmentation method is highly desirable. In the recent years it is found that the multi-channel Gabor decomposition represents an excellent tool for image segmentation and texture analysis. At higher frequency, Gabor filters have property to extract edge information. By analyzing such responses we have proposed a novel character segmentation method to segment top vowel modifier portion from an akshara (analogous to characters in English). Experiments are conducted on benchmark database of 1088 samples. Overall accuracy of 96.87% for top row index and 95.49% for consonant row index is observed.

KANNADA ALPHABETS RECOGNITION WITH APPLICATION TO BRAILLE TRANSLATION

In this paper, an effort is made to apply optical character recognition (OCR) for Braille translation on Kannada characters. In general, OCR systems for Indian language are more complex due to larger number of vowels, consonants, and conjuncts and Indian languages are inflectional and agglutinative in nature. Specifically, characters of Kannada script have higher similarity in shape and higher variability across fonts, making recognition of characters a difficult task. A decision tree is developed in this research work. The main motivations are that decision trees provide a natural way to incorporate prior knowledge of domain and many Kannada characters have similar looking shapes. The similar looking characters can be grouped and then further partitioned into categories at various levels to effectively create a decision tree. To facilitate this, three modular classifiers are developed based on the nature of Kannada characters. These modular classifiers are employed at nodes of the decision tree. The Braille equivalent of Kannada characters is obtained by translation rules. An overall accuracy of classification and Braille translation of 93.80% is obtained.

A Novel Method to Measure the Learning Capability of a Parameter in Artificial Neural Network with Application to Network Freezing

The artificial neural network is typically trained from initial weight/bias position. As training progresses the network parameters such as weights and biases are updated according to learning algorithm to reduce the performance index. Not all the network parameters are equally learning the input-output mapping. Some parameters would hold more discriminating capability while others are not so effective. We propose a novel method of measuring the learning capability of a network parameter. The learning capability for a parameter we call it as learnability is contribution of that parameter to reduce performance index as the network is training. The proposed method of measuring learnability is applied on network parameters freezing on feedforward neural network. Our method is validated on MNIST handwritten numeral database using backpropagation learning algorithm.