ASystem FOR ORIYA HANDWRITTEN NUMERAL RECOGNIZATION ...

2 downloads 0 Views 899KB Size Report
Some of these are post-card, inland letter, special envelope etc. ... Fig-2. Block Diagram of Handwritten numeral character recognisation. The first digit and.
International Journal of Applied Science & Technology Research Excellence Vol. 1, Issue 1, Nov-Dec 2011, ISSN NO. 2250 – 2718 (Print), 2250 – 2726 (Online)

A SYSTEM FOR ORIYA HANDWRITTEN NUMERAL RECOGNIZATION FOR INDIAN POSTAL AUTOMATION Manoj Kumar Mahto#1, Archana Kumari*2, S. C. Panigrahi#3 #

1

BRCM CET Bahal, Bhiwani, Haryana, India., * NERIST, Nirjuli, Itanagar, India, # B.I.T. Mesra, Ranchi India,

[email protected], [email protected], [email protected]

Abstract - In this paper, we present a system towards Indian postal automation based on PIN (Postal Index Number) code. Since India is a multilingual and multi-script country that was earlier colonized by UK, the address part may be written by combination of scripts such as Latin (English) and a local (state) script. Here, we shall consider Oriya script one of the local state language in India with English for recognition. It is very difficult to identify the script by which the PIN-code portion is written. So we have used two stage artificial neural network based general classifiers for the recognition of PIN-code digits written in Oriya. In this paper we propose a new technique namely, QuadrantMean technique to identify the numerals of PIN code written in Oriya script. By which the corresponding city name can be easily identified. The accuracy of the digit recognition module is 93.20%.

town/village. Both Table 1 and Fig 1 depict the mapping of the first two digits of the PIN-code to the corresponding regions of India. Recognition of handwritten numerals has a popular research area for many years because of its various application potentials. Some of its application areas are automatic postal sorting, bank cheque processing, form processing etc. Research on recognition of unconstrained handwritten numerals has made impressive progress in Roman, Chinese and Arabic scripts [8-14].

Keyword - Oriya Numerical Script, Postal Documents, Number recognition, Handwritten numerals, Neural Network.

1. INTRODUCTION Postal automation is a topic of research of interest for the two decades and many pieces of published article are available towards the automation of documents in non-Indian languages [1-7]. Several systems are also available for address reading in USA, UK, France, Canada and Australia. But no system is available for address reading of Indian postal documents. System development towards postal automation for a country like India may be more difficult in comparison to other countries because of its multi-lingual and multi-script behaviours. In India there are about 21 official languages and an Indian postal document may be written in any of these official languages. Moreover, some people write the destination address part of a postal document in two or more language scripts. Oriya is one of the popular languages in India. In India there is a wide variation in the types of postal documents. Some of these are post-card, inland letter, special envelope etc. Post-card, inland letter, special envelopes are sold from Indian post offices and there is a six-digit PIN code box to write pin number in a postal document. The PIN-code helps to locate a post office situated in a particular town/sub-

www.ijastre.org

Fig.1. The figure shows how the first digits of pin-code are mapped to different regions of India

Among various available approaches for this purpose, approach based on Neural Network(NN) has taken a considerable attention. Here the network architecture is, at first, trained by a set of training data and then the trained network is used to classify the input. Some researches suggest structural approach, where each pattern class is defined by structural description and the recognition process is performed according to structural similarities [8].

Page 17

International Journal of Applied Science & Technology Research Excellence Vol. 1, Issue 1, Nov-Dec 2011, ISSN NO. 2250 – 2718 (Print), 2250 – 2726 (Online)

TABLE 1 Representations of first and second digit of Indian pin-code.

The first digit and covering region First Region Digit

1 Northern 2

3 Western 4

5 6

Southern

7 Eastern 8

The First two digit and their representation First State/Circle covered two Digit 11 Delhi 12 to 13 14 to 16 17 18 to 19 20 to 26 27 to 28 30 to 34 36 to 39 40 to 44 45 to 48 49 50 to 53 56 to 59 60 to 64 67 to 69 70 to 74

Haryana Punjab Himachal Pradesh Jammu & Kashmir Uttar Pradesh Uttaranchal Rajasthan Gujarat Maharashtra Madhya Pradesh Chhattisgarh Andhra Pradesh Karnataka Tamil Nadu Kerala West Bengal

75 to 77 78 79 80 to 85

Orissa Assam North Eastern Bihar/Jharkhand

Another approach known as statistical approach, which is insensitive to pattern noise and distortion, can be used for numeral recognition [9]. However, modelling of statistical information is a tedious task. Among other available approaches are, support vector machines [10], Fourier and Wavelet description [11], Fuzzy rules [12], tolerant rough set [13], are reported in the literatures. Dutta et al. [16] used structural constraints such as the primitives imposed by the junctions present in the characters for recognition using a two -stage feed forward neural net. Pal et al. [17] proposed a structural feature based approach for Bangla handwritten numeral recognition. Bhattacharya et al. [18] used a modified Topology Adaptive Self-Organizing Neural Network to extract a vector skeleton from a binary numeral image. Multilayer Perceptrons (MLP) networks are then employed to recognize the same. Bhattacharya et al. [22] recently proposed a recognition scheme for Bangla handwritten numerals using multiple MLP classifiers and wavelet transform-based multiresolution pixel features. Roy et al. [25] discussed a system for Indian postal automation based on the pixel information of the normalized image. Basu et al. [19] used a two-pass

www.ijastre.org

approach for recognition. They used two NN classifiers; one for initial coarse classification and the other for final classification in the sub-groups obtained in the first pass. In this paper, we propose a new technique for Oriya handwritten numeral reorganization. The system takes a greylevel address image form the postal address document and segments it into lines and words. Then a coarse classification between the textual document and PIN code numerals is done through existing two stage artificial neural network based general classifiers. Each numeral of the PIN-code is then segmented and normalized. Next using the proposed, Quadrant-Mean technique PIN-code identification is done. The feature extraction is done based on the concept of splitting a numeral image into four quadrants and taking the mean of each pixel’s gray values form the respective quadrants. The merit of the proposed technique lies in its simplicity and accuracy of reorganization. At present, the accuracy of the digit recognition module is 93.20%. 2. PROPOSED TECHNIQUE There are many steps that need to be taken before reorganization of handwritten PIN-code numeral. The block diagram of the handwritten numeral character recognisation is shown in the Fig-2.

Character / Numeral acquisition

Pre-processing

Segmentation

Feature Extraction

Training of ANN

Testing and Result Fig-2. Block Diagram of Handwritten numeral character recognisation

Page 18

International Journal of Applied Science & Technology Research Excellence Vol. 1, Issue 1, Nov-Dec 2011, ISSN NO. 2250 – 2718 (Print), 2250 – 2726 (Online)

2.1. Character/Numerals Acquisition The first step in the recognition process is to acquire the handwritten numeral characters. Generally we use scanners to do job. Other sensor can be camera, video-camera etc. The input images are address blocks, shown in Fig-3, from live mail scanned at 300 pixels per inch with 256 shades of gray.

Fig-3. Examples of handwritten (in Oriya language) address blocks on postal documents

Data collection for the present work has been done from different individuals of various professionals and some of those are shown in Fig 3.b. Some numerals were also extracted from various postal documents. Since proposed technique to recognize numerals of Oriya language based upon structural shape, here it is worth to be mentioned that some of digits of Oriya language such as one, three, six, seven eight and nine are in practice have usually two structural shapes as shown in Fig-6. Due to this reason we have collected twice samples for the above digits as compare to the zero, two, four and five.

Fig-6. Sample of Oriya handwritten numerals

2.2. Pre-processing To get an idea of Oriya numerals and their variability in writing, a set of handwritten Oriya numerals are shown in Fig4.

Fig-4. Structural difference of Oriya numerals

(a)

(b)

(c)

The goal of pre-processing is to increase the quality of hand printed data. In pre-processing the preliminary steps include Segmentation, Normalization and Digitization. Here the segmentation is done in two steps. First a coarse classification between textual document and PIN-code numeral is done using two stage artificial network based general classifier. Secondly a fine classification is done to separate each numeral of the PIN-code. Since segmented digits/numerals vary in size, typically around 200 to 256 pixels, the numeral characters are then normalized using a linear transformation. This transformation preserves the aspect ratio of the character. Because of the linear transformation, the resulting image is not binary but has multiple gray levels, since a variable number of pixels in the original image can fall into a given pixel in the target image. The gray levels of each image are scaled and translated to fall within the range 0 to 1 without performing any skeletonization. Two sample of PIN code, taken from the address block of a live mail given in Fig-2, before and after normalization is given in Fig-7.

(a)

Fig.5. After binarization of sample numeric 5 in (a), 6 in (b) and 7 in (c).

www.ijastre.org

Page 19

International Journal of Applied Science & Technology Research Excellence Vol. 1, Issue 1, Nov-Dec 2011, ISSN NO. 2250 – 2718 (Print), 2250 – 2726 (Online)

2.4. Training of ANN (b) Fig-7. PIN code before and after normalization is shown in a and b respectively

2.3. Feature Extraction The feature selection is a very important criterion in handwritten numerals. They should be robust and easy to compute. Here we have considered the mean of the object pixels as one of the important feature to identify the required numerals. But whenever we are taking mean of the required numerals there might have possibilities that the mean two numeral may coincide. Moreover, the problem is also persists when we split the numeral in two different quadrant and taking the mean of each pixel’s gray values form the respective quadrants. The illustration of the above situations in case of one and two is shown in the Fig-8.

Feedforward, multilayered neural network is used for the recognition of numeral characters. An Error Back-propagation (EBP) algorithm is used to train the neural network. Error at the output neurons are computed by considering the differences between the actual output of each neuron and its desired value as given for the input pattern taken from the training set. The mathematical expression for the error, E, computed with the output neurons for some training is given as

E

1 2 (d j  O j )  2 j

(2)

Where dj and Oj respectively be the desired and actual output of the jth output neurons for some input vector. BP algorithm, in a nutshell, minimizes the sum of squared errors for training patterns by conducting a gradient descent search in the weight space. So the amount of change, ∂Wji, that is to be made to the weight value of the connection from the output of the ith to the input of the jth neuron, is given as

Fig-8. Feature extraction of numeric

W ji  

Hence for simplicity and better performance in the proposed Quadrant-Mean Technique we have spited the image matrix of the numeral characters, say of order n  n, into four sub matrices of order n/2  n/2 as shown in the Fig-9. Then the density/mean of the object pixels is calculated from the respective sub matrices correspondence to each of the four quadrants using the formula

i 

li mi

, i = 1, 2, 3, 4

(1)

where li and mi respectively be the number of object pixels and number of total pixels in the ith quadrant, 1st Quad

4th Quad

2nd Quad

3rd Quad

E W ji

.....(3)

Where η (0< η