International Journal of Computer Applications (0975 – 8887) Volume 146 – No.3, July 2016 29 Number Recognition from Captured Images Randhir J. Patil Assistant Professor Sanjay Ghodawat Institutes, Atigre, India Akshay G. Bhosale Assistant Professor Sanjay Ghodawat Institutes, Atigre, India ABSTRACT Now-a-days because of digitization, it is very important to have all the data in the form of soft copy so that it is easy to store and maintain. Often the handwritten documents are available, but again if we want to use the data written in that document then we need to type that data in any word processing software and then use it. This is very time consuming and tedious job. In this paper an attempt is made to automatically recognize numbers in the document by capturing images of that document. Till now emphasis is given only on recognizing digits or numbers from images. This methodology is implemented in automatic marks filling system. This system captures images of the tables in which marks are written by the teacher in the students‟ answer book, recognize the numbers or marks from it and fill the same in appropriate database on the computer. The fundamental technique used here is Dynamic Time Warping (DTW). Although this technique is used for speech processing, this paper describes its exclusive use in number recognition from its images. The system is implemented and simulated in MATLAB. General Terms Pattern Recognition, Character Matching Keywords Image Processing, DTW Technique, Data Acquisition, Data Transformation, Number Recognition, Automatic Marks filling System. 1. INTRODUCTION Due to digitization it is necessary that whatever data we have that must be available in the form of soft copy. If the data is available in the form of soft copy then it is very easy to maintain it, store it and also process it. But if the data is available in the form of hard copy or handwritten form then it is very time consuming and tedious job to convert the data available in hard format to soft format. It is generally accepted that handwritten characters are hard to recognize. The primary reason is that characters are written in uneven manner. There are some electronic notepads which can be attached to a computer, so whatever we write on that notepad gets converted to digital text. This requires additional software to be installed on the computer. But if we have data already written on a piece of paper then it is difficult to directly convert it to digital format. In such cases this system can work well. There are several technique for isolated template recognition, like Optical Character Recognition (OCR), Artificial Neural Network (ANN), Correlation, etc. Several investigations have been carried out to determine the best technique. It is necessary to develop a system which may have a complex recognition algorithm and may have lower accuracy. But if limited characters are considered like 0 to 9 numbers only, then there is possibility to increase accuracy as well. The DTW technique that has been described here is most probably used in speech template recognition. It is also applicable to the isolated speech recognition. In speech processing two feature extracted matrices of respective speech templates are compared to find minimum distance. Same technique is used here to develop this system, but two matrices are generated from captured images of characters. The ultimate objective here is to compare two matrices generated from images of the number. The rest of the paper is organized in different sections as follows. Section II describes the basics of DTW technique. Section III describes detail methodology and the algorithm used in implemented system. After implementing this system in MATLAB, results were obtained and those are discussed in Section IV. Finally, Section V describes the future scope concluding remarks. 2. DTW ALGORITHM The key idea behind DTW is to align two non-temporal speech sequences and finding minimal DTW score between them. Consider following example shown in figure 1. This figure illustrates the DTW technique used in speech processing. In DTW matrix is computed using MFCC coefficient values. Take all the frames (each frame consist of 13 MFCC coefficient) of stored voice symptom on y-axis and that of testing symptom on x-axis [1][2][3][4]. Fig 1: DTW aligning Calculate local distance between two corresponding frames using Euclidian distance formula sqrt [(X i - Y j ) 2 ]. Thus cell (n, m) of DTW matrix is result of Euclidian distance between nth frame of test sample and m th frame of reference sample [5][6][7][8]. Then minimum DTW path is constructed through cell (1,1) to cell(r, t) where „r‟ and „t‟ are last frames of reference and test sample respectively. Every next cell is chosen using formula min (celli, j celli-1,j celli,j-1). Thus by calculating DTW path between test signal and all reference signal, nearest neighborhood approach gives recognized word/symptom [9][10]. Here instead of speech features