Optical Character Recognition System - Assignment Example

Name Optical character recognition system Tutor Institution Date Introduction Optical character recognition system is kind of software that was engineered to convert typewritten or hand-written text documents into machine text that can be formatted. Entering data through optical character recognition is quite fast, more efficient and more accurate. In most cases of conversion of text, the OCR offers a substitute to keyboarding or data entry. Before using the optical character recognition, the source material is first scanned specifically using an optical scanner but sometimes a particular circuit board in the PC. The purpose of the scan is to read the source material as a pattern of dots or rather as a bitmap. Software that recognizes the images is also needed. The OCR software afterwards processes the scans so as to differentiate between text and images and also establish what letters are represented in the dark as well as light areas (Lais, 2000). Image Segmentation Image segmentation is a pizel classification process that is purposed to segment or extract regions or objects from the background. It described as a significant preprocessing step to the ultimate success of image recognition, image compression, image visualization and image retrieval. One should note that there is no set sequence to the approach to segmentation. Various types of parts of a scene can act as the segments upon which descriptions are based on and there are a variety of ways a person can utilize to extract the parts from the image. Appropriate selection of segmentation technique is based on the type of applications and images. In optical character recognition, the text is detached from the document image and further divided into lines, words, columns, and connected constituents. In the process of building character sub-images, a person is confronted with broken or touching characters that appear in degraded documents like scan, fax and photocopy. The evaluation that has been made concerning segmentation algorithms is subjective and hence people make their judgment based on the outcome of several segmented images or according to their intuition. However, the segmentation algorithms can be evaluated through the use of symmetric along with asymmetric distance metric alternatives (Shih, 2009). In today’s OCR engines, the technology multiple algorithms of neutral network is added to examine the discontinuity line between the text characters, the stroke edge and the background. With the presence of irregularities of having printed ink on paper, every algorithm then makes an average of the dark and light along the stroke side and through matching it to known character it makes a smart guess concerning what character it is. Further, the OCR software polls the results or finds the average from the entire algorithm to get a single reading (Lais, 2002). Methods of segmentation Thresholding Thresholding is one of the different methods of segmentation. This method is based on the hypothesis that clusters in the histogram correlate to either objects or background of interest that can be got by separating the histogram clusters. This is a process where the gray-scale or color- image is minimized to a binary image by use of an optimal threshold. The purpose is to extract the pixels from an image which characterize an object such as a line image like a map or graph or a text image. Even though the information is represented in binary form, the pixels represent a wide range of intensities. Therefore the purpose of binarization is to mark those pixels that belong to the background regions with diverse intensities and true foreground regions with a definite intensity (Devi, 2006)). An effective thresholding algorithm must preserve semantic and logical content. There are two forms of thresholding algorithms namely adaptive or local thresholding algorithms and global thresholding algorithms. In local thresholding, diverse threshold values for various local areas are applied. In global thresholding, a particular threshold for each and every image pixels is used (Devi, 2006). It is not possible to segment all images successfully into background and foreground when simple thresholding is used. In cases where the intensity distribution of objects in both the foreground and background is different, this method is quite applicable. In adaptive thresholding, various threshold values for various local areas are applied. There are diverse methods of selecting the threshold value (Shih, 2009). Edge detection In edge detection method, the edges classify boundaries and the regions are enclosed within these edges. Edge detection is an important tool in computer vision especially in areas of feature extraction and feature detection. The purpose is to identify the regions where the brightness of the image changes sharply in a digital image. Determining the position of the edges is the first step in segmenting images into regions (Lais, 2002). When edge detector is applied in an image, the result is a set of connected curves that show the boundaries of objects and the boundaries of surface markings. It means that when edge detection algorithms is used, the amount of data purposed to be processed in the image may be reduced hence filtering out information regarded as irrelevant and at the same time preserving the significant structural properties of an image (Shih, 2009). Split and merge operations Segmentation algorithms that employ the split and merge operations are basically restricted to a split stage which is consequently followed by a merge stage. The pure merging methods are computationally expensive since they begin from such small individual points. However, this can be made efficient by recursively dividing the image into much smaller and smaller regions until the point where single regions are coherent and then merge the images recursively to come up with larger coherent regions. The first step therefore is splitting the image where the entire image is considered as one region. In cases where the entire region is coherent that is all the pixels represented in the region are similar, then the region is left unmodified. However, where the region is not adequately coherent, splitting must be done (Govindaraju & Setlur, 2009). Feature extraction Selecting the feature extraction method is basically the most fundamental factor in the achievement of high recognition pattern. Feature extraction is defined as problem of extracting the most relevant information from raw data for classification purposes. The sense is to enhance the between -class pattern variability while minimizing the pattern within- class. It should be noted that various feature extraction methods conform to this requirement to certain degree, depending on the available data and particular recognition problem (Jain, 1996). In OCR, a feature extraction method may prove to be successful in a single application domain but fail to work successfully in another application domain (Yampolskiy, 2007). A good feature extraction method should be that which give the best selection method for a particular application. It is wise to consider whether the characters under scrutiny have known size and orientation, whether they are typed, machine-printed or handwritten and to what level they are degraded. Some methods of feature extraction work best on gray-level sub-images of particular characters while other do best on solid four-or- eight- attached symbols divided from the binary raster image, thinned symbols or symbol contours. Furthermore, the format of the extracted features should match with the requirements of the selected classifier. Grammar or graph based descriptions of the characters are best for syntactic or structural classifier. On the other hand discrete features that may take two or three different values are well suited for decision trees (Jain, 1996). Pattern recognition is a discipline whose main goal is classifying objects into various classes or categories. Depending on the application, the objects can either be signal waveforms or images or any other form of measurement that require to be classified. Character recognition is a very significant area of pattern recognition. The optical character recognition basically has a front end device which is composed of a scan lens, a light source, a detector and a document transport. The variation of light intensity is translated into numbers at the output of the light sensitive detector and hence an image array is formed. In the process, a series of image processing procedures are applied resulting to line and character segmentation (Govindaraju & Setlur, 2009). Next, the pattern recognition software then proceeds to recognize the characters a process which involves classifying each and every character in its correct number, letter and punctuation class. There is an advantage of storing the recognized document rather than the scanned image. There is a great interest invested in systems that identify handwriting besides the printed character recognition systems. A very good example of an application of such systems is in the machine that reads bank checks. It is a requirement that the machine recognizes the amount stated in digits and figures and match them appropriately. In addition, the machine could check whether the account to be credited corresponds to the payee. Much labor could be saved by the machine even though only half of the checks are checked correctly. A similar application is in the postal code identification where automatic mail-sorting machines are used in the post offices (Jain, 1996). Classification Image classification explores the numerical properties of different image features and further organizes data into groups. One mode of classification in optical character recognition is classification of the blocks which is a form of text and non-text block reconstruction. Spatial domain features opens a door way for good block segmentation of grayscale postal images. The classification error is reduced when first pass classification is combined with context-based classification. The document classification trials automate the process of recognizing, sorting as well as routing of documents to their right destination (Shih, 2009). Conclusion The current OCR software is able to recognize source material with different fonts. However, the software experiences problems in recognizing script and handwriting fonts that imitate handwriting. This will not be a problem soon as developers are considering various approaches that will improve handwriting as well as script recognition. Advances have made the system reliable as a minimum of 90% accuracy is expected for average quality documents. The initial step in achieving better recognition starts with the scanner. The quality of the scanner’s light arrays device ultimately affects the OCR results. The more tightly the arrays are packed, the more the scanner can detect more distinct colors and finer image. Background or smudges color can make a fool to the recognition software and therefore to improve the recognition rate or refine the image the scan resolution can be adjusted. Quadratic Integral Ratio (QIR) algorithm (Global threshold) Three sub images of QIR method References Devi, A. 2006. Thresholding: A Pixel-Level Image Processing Methodology Preprocessing Technique for an OCR System for the Brahmi Script. http://www.ancient-asia- journal.com/index.php/aa/article/viewArticle/13/25 Retrieved 21st April, 2011 Govindaraju, V.& Setlur, S. 2009. Guide to OCR for Indic Scripts: Document Recognition and Retrieval. New York, Splinger. Jain, A.1996 FEATURE EXTRACTION METHODS FOR CHARACTER RECOGNITION--A SURVEYhttp://www.cs.sfu.ca/cc/414/li/material/refs/OCR-survey- 96.pdf Retrieved 21st April, 2011 Lais, S. 2002. QuickStudy: Optical Character Recognition http://www.computerworld.com/s/article/73023/Optical_Character_Recognition July 29, 2002 Retrieved 21st April, 2011 Shih, F. 2009. Image Processing and Pattern Recognition: Fundamentals and Techniques. New York,Wiley-IEEE Yampolskiy, R. 2007. Feature Extraction Approaches for Optical Character Recognition. Briviba Scientific Press, Read More