WORD RECOGNITION OF KANNADA TEXT IN SCENE IMAGES USING NEURAL NETWORK.

Puneet Shetteppanavar and Aravinda Dara. Department of Computer Science and Engineering, Sree Dattha Institute of Engineering and Science, Ibrahimpatnam, Hyderabad, 501510, India. ...................................................................................................................... Manuscript Info Abstract ......................... ........................................................................ Manuscript History


ISSN: 2320-5407
Int. J. Adv. Res. 5 (11), 1007-1016 1008 characters. Natural scene images have numerous and various degradations and to handle these issues, it is very important to link consecutive steps together to take dynamic decisions.
The proposed approach is more closely related to those in [4,6,7], which address the problem of word recognition of Kannada text in scene images using neural network. On one hand, these methods recognize text with a significant accuracy, but on the other hand, their recognition results leave a lot to be desired. Note that the challenges of this task are evident from the best published accuracy of only 85.53% on the scene text dataset [4]. The probabilistic approach we propose in this paper achieves an accuracy of over 97.17% under identical experimental settings.
The remainder of the paper is organized as follows. In Section 2 we present our proposed methodology. We provide results on two public datasets and compare our method to related work in Section 3. Implementation details are also given in this section.

PROPOSED METHODOLOGY:-
The proposed method uses zone wise horizontal and vertical profile based features for word recognition. The method consists of various phases such as, pre-processing and segmentation, feature extraction, training neural network and word recognition model. Figure 1 shows the block diagram of the proposed model. The detailed description of each is given in the following subsections.

A. Pre-processing and Segmentation
The scene text images have issues like lighting effects, shadowing, blur, color degradation and size etc. The purpose of this phase is to make the images to be of standard size and remove complex backgrounds easier for further processing. Pre-processing and segmentation procedure consists of several steps, which are as detailed below;

Binarization
The input word image is converted into binary image that has only two possible values for each pixel represented by either 0 or 1. Every word image is resized to fixed size based on length.

Segmentation
In this phase, the word image is portioned into individual character images. The goal of segmentation is to simplify an image that is more meaningful and easier to analyze. This segmentation step is crucial for the word recognition: any error directly reduces the recognition accuracy of the system.

Thinning
Thinning refers to the process of reducing the width of a line like object from many pixels wide to just single pixel. This process can remove irregularities in letters and in turn, makes the recognition algorithm simpler because they only have to operate on a character stroke, which is only one pixel wide.

1009
Bounding Box Generation Before analysing any character of the word image, it is important to identify the (pixel) boundaries of that character. Thus, a bounding box is generated around the image.

B. Feature extraction
Features are extracted from the segmented pre-processed characters of the word image (i.e. word image may be 2,3,4,5 or 6 character image). Each character image is divided into 6 vertical zones and 6 horizontal zones, where size of each horizontal zone is 5x30 and the size of each vertical zone is 30x5. Then sum of all pixels in every zone is determined as a feature value. Totally 12 features are obtained from each character stored in feature vector. Feature vector is shown in the equation (1).
= [ ] (1) The feature vector for word image consisting of "N" number of characters is depicted in equation (2).
(2) Where 1<=i<=N and 2<=N<=6 is the Feature Vector of the ith character of the word image. is the Word feature vector comprising feature of all characters.

C. Neural Network Training
The Feed Forward Neural Network is used for training the network. The features generated from the training database are used to train the model. The Neural Network that is used for the training has 36 input neurons, 38 hidden neurons and 6 output neurons. Figure 2 shows the Neural Network model used for training. Test image is processed to obtain zone wise horizontal and vertical profile based features, which are further fed to Neural Network for recognition.

EXPERIMENTAL RESULTS AND DISCUSSIONS:-
The dataset is collected from government office display boards, traffic boards and boards on various buildings in Karnataka. The dataset consists, 750 images of Kannada words. The proposed methodology for word recognition system has been evaluated for several samples dealing with various issues. The method achieves recognition accuracy of 97.17%. The system is efficient and insensitive to the variations in size and font, noise, blur and other degradations.

A. Sample Kannada Word Image containing Blur
The word image that is selected from the database is given in figure 3. The image has several challenges like unusual fonts and size, blur etc. Pre-processing step is performed to make the images to be of standard size, remove complex backgrounds and makes them easier for further processing in word recognition. In Pre-processing step the color image is converted into gray scale image, then into binary image. The binary image is resized and then applying thinning process and bounding box. Then features for each word are extracted and then testing is performed.

B. Sample Kannada Word Image with Variable Font Size
The word image that is selected from the database is given in figure 4. The image has challenge like variable font size. The resulting gray scale image is shown in figure 4.1.      21 20  22  15  24  36  16 7  46  18  10  41  23 18  23  23  16  13  5 32  20  20  21  18 The method is able recognize word correctly in the presence of variable font size. The recognized pattern is given in table 2.

C. Sample Kannada Word Image with Dark Background
The word image that is selected from the database is given in figure 5. The image has challenge like dark background. The resulting gray scale image and resized binary image is shown in figure 5.1 and figure 5.2 respectively.      36 49  30  34  42  36  57 27  35  38  44  26  22 16  23  16  28  24  5 10 32 23 32 27 The method is able recognize word correctly in the presence of dark background. The recognized pattern is given in table 5.     29  24  19  37  15  23  5  23  28  15  23  53  12  20  26  19  20  22  5  26  27  31  6 24 The method is able recognize word correctly in the presence of dark background. The recognized pattern is given in table 6.

CONCLUSION:-
This work strives toward a novel methodology that aids pre-processing, segmentation and recognition of Kannada Words from camera based images. The proposed methodology is based on zone wise horizontal and vertical profile based features and neural network as a classifier for Kannada Word Recognition. The system works in two phases training phase and testing phase. Exhaustive experiments are done for analysis of zone wise horizontal and vertical profile based features.
The system successfully processes camera based images having challenges like variable lightning condition, noise, blur, unusual fonts etc. The methodology is tested with 750 samples and gives recognition accuracy of 97.17%. The method can be extended for word recognition considering new set of features and classification algorithms.