http://dlib.iit.ac.lk/xmlui/handle/123456789/274 Tue, 07 Apr 2026 08:42:34 GMT 2026-04-07T08:42:34Z http://dlib.iit.ac.lk/xmlui/handle/123456789/484 Instance segmentation based objects detection in digital documents Balasubramaniam, Bravin Digital documents have increased in numbers exponentially within the last twenty years. Because of this information captured in digital documents also lost vastly. There are multiple researches done on using Natural Language Processing to mechanically extracting, understanding and, eventually, summarizing key data from digital documents. However, while text is without argument, a basic way to convey data, there are contexts where graphical components are far more powerful. For example, in scientific research papers, several experiments, variables and numbers must be reported in a concise manner that fits better with tables/figures than text. Graphical components possess in conveying information that may be otherwise cumbersome to explain in words, each for the author to express and also the reader to grasp. We developed an application which can identify/detect any graphical components in given digital document and extract them separately. Application not only have the capability to extract these graphical components but it also can classify them into three different categories/classes. 1. Tables 2. Charts 3. Other (Any other graphical components other than tables and charts) The results shows that graphical components are extracted from digital documents and classified correctly with an 81% of accuracy Tue, 01 Jan 2019 00:00:00 GMT http://dlib.iit.ac.lk/xmlui/handle/123456789/484 2019-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/483 Classification of Tender Notices through Deep Learning Concepts Vanniasinkam, Damian Perrin With the continuous growth of the tender industry in Sri Lanka, the visibility of tenders to the suppliers is a cause for concern. Tender alert service companies have minimised this complexity but due to the cumbersome task of sourcing papers daily and categorising tenders, they are not always efficient or accurate. Keeping in mind the escalated growth of the machine learning industry, this thesis takes the tender alert service industry in to consideration and looks at multi-label classification methods to automatically classify tenders. The goal is to identify what categories a tender belongs to by just considering its heading. A dataset with 36 tender categories was used for this purpose. A thorough literature review session was conducted to identify the best methods to classify the tenders and it was identified that converting the multi-label problem to binary problems was the best solution to ensure high accuracy. As such, the data was processed through 3 classifiers, namely, Linear SVM, Random Forest and Neural Networks, to identify suitable classifiers for each category. Apart from a single category that didn’t have many supporting records, all other categories were able to produce classifiers with a f1-score of above 85%. Although the hamming loss could have been better, most of them were below 10%. Based on this, we can conclude that a satisfactory classification model for tenders was achieved. Tue, 01 Jan 2019 00:00:00 GMT http://dlib.iit.ac.lk/xmlui/handle/123456789/483 2019-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/482 ‘Affect-Pulse’ Intelligent Solution for Neuro Marketing using Machine Learning Techniques Soris, Trevor Fabian Marketing plays a significant role in any kinds of business. Product marketing requires Indepth researching, advertising, and selling products or services to reach out to the target audience. Retailing is undergoing rapid growth with the rising dominance of e-commerce. Around the world, e-commerce is changing the way people shop around the world. To withstand the emerging flow in e-commerce, the marketer has to ensure that the correct product reaches the target audience at the correct time. When it comes to e-commerce websites, visual content can make or break a website. It plays a pivotal role in making a successful sale or putting off a potential customer. Current statistics on e-commerce suggest that 22% of the total sales were returned since the product received did not look the same. Visual content is the only communicating medium about the product. It is essential to make sure that the target audience will like the content posted in the ecommerce page emotions is one of the ways to communicate or express their self. A person's emotions can be identified from their gesture, facial movements, and Body language. However, finding the emotions from these methods can get quite tricky.. In this research, the proposed solution will aid the marketers of the product, to identify/predict likeness towards a visual medium before it goes online in their respective e-commerce sites Human emotion detection using Electroencephalogram (EEG) started playing a crucial role in developing a smarter Brain-Computer Interface (BCI). In this research, DEAP physiological, emotional database is used to identify emotions using the arousal valance model. Using wavelet transformation, the EEG signal was decomposed into four frequency bands (theta, alpha beta, and gamma). The Daubechies order 4 wavelet function (db4) was utilized to do the processing. From these frequency bands, linear (Energy, Power Spectral Density) and nonlinear (Entropy) statistical features were extracted, which were used in the machine learning classification algorithms. Many experiments were carried out on different types of classification algorithms such Neural Network , Support Vector Machines and Bayesian Networks . From the experiments it was found Neural network produced the most accurate models. Once the model built, using the DEAP dataset, knowledge learned was transferred to another new domain, this research it was Neurosky Mindwave Mobile device to classify the emotions from the RAW EEG signals. Finally, the prototype was evaluated by the potential users, and some future enhancements were identified Tue, 01 Jan 2019 00:00:00 GMT http://dlib.iit.ac.lk/xmlui/handle/123456789/482 2019-01-01T00:00:00Z http://dlib.iit.ac.lk/xmlui/handle/123456789/481 Detection of Hot Topics on Twitter using Named Entities and Event based Incremental Clustering Jayasumana, Mallika Arachchige Prasad Akilendra Social media a place where most of the people spend their day to day lives in. It is a place where people communicate and interact with each other, a place where they share their events and get updated on current events and a place where a lot of people are active most of the time. Many parties will largely value from identifying the current trending hot topics as it will help their business. Marketing companies using trending hashtags when marketing their products are more likely to be noticed. News companies will be able to find the latest news and relevant feedback for them. While there are many approaches to detect trending topics most of the existing systems have not given much thought to real time performance and have failed to remove unnecessary noise in data making them inefficient. This research investigates on how to extract events from a twitter stream of data in real time and display them in the form of hot topics. To achieve this an incremental event clustering approach is taken which would be based on the named entities of the tweets. The use of pretrained Doc2Vec generated vectors was proposed to be used for clustering the tweets into their respective events. Additionally, the tweets will undergo a pre-processing stage where noise is removed and an event merging process where similar tweets are merged to the same cluster. After testing and evaluation phase, the implemented DOH framework gave a Normalised Mutual Information score of 0.911 and a Rand Index of 0.794 after testing it on 100 labelled tweets. The proposed methods and algorithm have proven feasible and given successful results this is justified. Tue, 01 Jan 2019 00:00:00 GMT http://dlib.iit.ac.lk/xmlui/handle/123456789/481 2019-01-01T00:00:00Z