IRSPAS 2018
Permanent URI for this collectionhttp://repository.kln.ac.lk/handle/123456789/19084
Browse
3 results
Search Results
Item Sinhala handwritten address recognition for postal sorting automation(Research Symposium on Pure and Applied Sciences, 2018 Faculty of Science, University of Kelaniya, Sri Lanka, 2018) Ifhaam, M. F. A.; Jayalal, S. G. V. S.Sri Lankan post office mail sorting process is done manually, even today. Even though employees are well experienced, it takes considerable time and pushes employees to work overtime in places like Central Mail Exchange (CME). With major issues like unclear handwriting, having trouble to recognize some uncommon or ambiguous names and carrying these duties twice a day create a negative impact on the efficiency of the postal delivery system. In the prevailing system, forward mails and delivery mails are the two categories of separating mails at the sorting centers. Delivery mails are the posts which can be delivered to its destination directly. Forward mails are the ones which needs to be sent to an appropriate post office that can deliver the particular post to its destination. Majority of Sri Lankans use Sinhala language for their day to day activities. Then again, less researches had been done on Sinhala handwriting recognition. However, the research recognized only the city names by using the postal sorting domain in 2004. This finding is useful for main post offices, but the number of main post offices in Sri Lanka are limited to 501 whereas, number of sub post offices in Sri Lanka are mentioned as 2953. Sorting process at sub post offices needs street address or one line below in the address line, which cannot be done only using city names. These requirements made the above research less successful. Though similar systems have been implemented in other languages like English, there haven't been a highlighting findings except the above mentioned research due to the complexity of languages like Sinhala. The proposed system is focused on recognition of Sinhala handwriting using Optical Character Recognition (OCR) and image processing technologies. Implementing these techniques to recognize postal address will increase the efficiency of postal mail sorting. Handwritten postal envelopes will be used as the training, testing and validating materials. Therefore, this research is not limited only to a single restricted writing style, but also for unstructured writing styles. In this system, one of the major impediment is touching characters. Segmentation of handwritten touching characters become a crucial step in such systems. Conventional segmentation methods are incapable of handling the complexity of Sinhala handwriting. The proposed method separates touching characters into isolated character models in two steps as described in literature such as basic projection profile method and water reservoir concept. Recognition quality will be extended using heuristics since population of recognizing words are finite. Genetic Algorithms (GA) will be used to generate more optimized results faster with higher accuracy. The Primary Objective of the research would to identify the automatic way of forwarding the letter to the next post office from the current post office. Given addresses are written in the default format. This format can be extended to more formats as improvements in future. Since current methods are completely manual, evaluation should be done with the help of experienced employees at sorting centers.Item Applicability of machine learning techniques to improve the accuracy of communication of children with isolated speech and language delay in Sri Lanka(Research Symposium on Pure and Applied Sciences, 2018 Faculty of Science, University of Kelaniya, Sri Lanka, 2018) Dissanayake, D. M. N. M.; Jayalal, S. G. V. S.Children with speech and language difficulties face a lot of inconvenience when they try to communicate with others. This study looks for the possibility to help improve their communication by integrating Machine Learning (ML) techniques to widely used picturebased communication method. Augmentative and Alternative Communication (AAC) is a popular intervention used for the treatment of speech delay, which includes exchangebased pictorial communication. Apart from the manual implementations, the mobile applications based on AAC have become popular among the interested audience, where they can place the relevant pictures instead of typing words and the pictures are spoken out by the application. The issue with the existing solutions was that the application just reads out the referents of each picture without making a proper sentence, sometimes making it hard to understand what the patient has meant, along with less support provided for local language. Main objective of this study was to explore the possibility of integrating ML techniques to the AAC based mobile application, which would be ordering the pictures placed by the patient better and predict better picture-based suggestions, increasing the accuracy of communication. In the implementation, a text classifier based on Natural Language Processing (NLP), which is a ML technique, was used to assign a class to each chosen referent. The NLP model was trained using a labelled dataset which contained referents and the labels they belong. Then an algorithm was written to reorder the pictures placed, using the referents and assigned classes. A sample of 12 children diagnosed with isolated speech and language delay was used to test the application. They were tested twice; once with the normal application and again with the enhanced application. The application was used to communicate with both their regular and non-regular communication partners. Majority out of the 12 partners were positive on the improved accuracy of the communication with the enhanced application. Previously used similar applications had not used ML techniques to enhance the accuracy of the output of the application. Categorizing the pictures had been already done; yet, new data had to be repeatedly added to the categories manually, and at another level, a meaningful sentence was not formed as the output. Findings of this research proved that integrating Machine learning techniques such as NLP, to order the output of the application making more sense, was successful in terms of accuracy and the meaning of communication.Item Prediction of type 2 diabetes risk factor using machine learning in Sri Lanka(Research Symposium on Pure and Applied Sciences, 2018 Faculty of Science, University of Kelaniya, Sri Lanka, 2018) Menike, R. M. S. D.; Jayalal, S. G. V. S.; Algiriyage, N.Diabetes mellitus is in third place in the index of 20 major diseases affecting deaths in Sri Lanka. Diagnosis of diabetes is a key and insipid task. A successful, easy and correct method has not been identified to identify the diabetes mellitus in the early stage. Currently, the Diabetes detection is done using blood tests, such as Glycated hemoglobin (A1C) test, Random blood sugar test, Fasting Plasma Glucose test, Oral Glucose Tolerance Test, and Blood Sugar Series. People who do not have a special condition are generally unwilling to go for a blood test, which is a process that costs them time and money. Diabetes mellitus cannot be fully cured, but if identified in prediabetes, it is possible to prevent prediabetes from developing into type II by the actions such as eating healthy foods, losing weight, being physically active. As there are no regular medical checkups to diagnose pre-diabetes among the general public, identification of pre-diabetes is problematic in Sri Lanka. Machine learning techniques have been successfully applied to predict the risk factor for diabetes mellitus in other countries. Due to the high variance of economic and cultural factors, it is very difficult to come up with a common model to all countries. The detection of diabetes from some important risk factors is a multi-layered process. This research is primarily aimed at identifying factors that contribute to the prevalence of diabetes in Sri Lanka and finding a mechanism to predict the risk of diabetes through the use of machine learning algorithms over the identified factors. The gathered dataset consists of anthropometric and behavioral data of a set of people who have diabetes and don’t have, such as age, BMI, gender, heredity, and Hypertension etc. Wrapper methods are used to identify the most influential factors of these factors that affect diabetes mellitus. Since earlier studies have shown better performances, Support Vector Machine, J48, Random Forest algorithms are used for classification of selected dataset. As a result, three models are generated, and the performance of each model is measured analyzing measurements such as accuracy, specificity sensitivity. The outcome model of the study is the one that shows the best performance. That model presented by this scrutiny as the final output can be used by the public without specific domain knowledge, provide a more accurate clue of the diabetes risk of themselves by giving the data related to the identified factors as inputs.