Handwritten Off-line Recognition of Names (HORN)Dafna BittonDepartment of Computer Science and EngineeringUniversity of California, San DiegoLa Jolla, CA [email protected] sizes at universities can be in the hundreds. The process of returning gradedwork to students can be time consuming and tedious. In order to avoid this process,we propose a method to return work to students online. This process providesfeedback to students when it is available and avoids manually returning papers inclass. The course staff will scan in all of the graded homework and tests, which canbe automated, and label each document with the corresponding student’s name.The laborious part of this process is labeling the documents with the appropriatename. This project addresses this problem by proposing a method to automate thisprocess by matching the name on the document with a name on the class roster.1 IntroductionI took CSE 190a 1 year ago, in winter quarter of 2008. This quarter I will continue the project Istarted then. At the end of winter 2008, I got to the point where I could segment characters fromthe input document and run the nearest neighbor algorithm on them against a fairly small amountof training data taken from the ABCDETC dataset. The level of accuracy was low but the projectwas left in a good place to build upon it. This quarter I plan to apply different machine learningalgorithms to the problem, such as boosting, support vector machines and logistic regression. Iintend to ditch the HMM technique used last year, but may eventually incorporate it to the model.Two things to keep in mind, that have not changed from last year:• I assume that I have a complete roster of the class, and will use this information in the finalclassifier.• I will be using character boxes to assist in image segmentation.2 QualificationsI am a 1st year master’s student in the computer science department. I have taken the followingrelevant courses at UCSD:Course Title Instructor quarterCSE 166 Image Processing Belongie Fall 2007CSE 250A Probabilistic Reasoning and Decision-Making Saul Fall 2007CSE 151 Introduction to Artificial Intelligence: Learning Elkan Fall 2007CSE 190A Projects in Vision and Learning Belongie Winter 2008CSE 291 Web-Scale Information Retrieval and Data Mining Elkan Spring 20083 Milestones• Weeks 1,2,3 - Research alternative machine learning techniques to apply to the problem.Either find existing Matlab implementations for these techniques or implement them my-self. Research additional sources of training data. Alternatively, have 10 to 20 people fillout empty ABCDETC sheets, scan them in, and run my code from last year to isolate thecharacters to create more training data.• Weeks 3,4,5 - Apply the techniques found in previous weeks to the problem and find whatworks best.• Weeks 6,7,8,9 - Based on previous weeks, make any adjustments necessary. Experimentwith different filters, jittering training data, combining multiple machine learning tech-niques to finalize the system.4 Questions1. Which feature extraction method produces the highest accuracy?2. How much training data is necessary?3. What is the highest level of accuracy achievable?4. Is this a realistic problem to solve in a quarter (or two)?5 Use of Existing Software and DatasetsI will be building upon my code from last year and again using the ABCDETC datasets as trainingdata to begin with. All code will be written in Matlab.6 References[1] Trier, O.D., Jain, A.K., Taxt, T. (1995) Feature Extraction Methods For Character Recognition – A Survey.[2] Edwards, J., Forsyth, D. (2005) Searching for Character Models.[3] Bitton, D. (2008) Recognition of Handwritten
View Full Document
Unlocking...