Browsing by Author "Mehendran, Y."

Now showing 1 - 4 of 4
  • No Thumbnail Available
    Item
    Analysis of Traffic Sign Detection and Recognition Techniques
    (Uva Wellassa University of Sri Lanka, 2021) Shanmugam, S.; Tharmaratnam, B.; Sandradeva, T.; Mehendran, Y.
    Automated Traffic Sign Detection and Recognition (ATSDR) is a trending research field in this current decade. It is a very important part of the intelligent transportation system as traffic signs assist the drivers to drive more carefully. This paper provides a review of three major steps in the ATSDR system; video segmentation, detection, and recognition. There are many techniques used for the detection and recognition process. However, those techniques are affected by different internal and external conditions like camera quality(fps), lighting conditions, time periods, etc. The main objectives are; to identify the different traffic sign detection and recognition techniques, develop the ATSDR system by using those selected technologies and analyze the performance of those techniques in different lighting conditions and time periods in Sri Lanka. Real time video sequences of traffic signs were collected and partitioned into single frames using video segmentation. The traffic signs were detected using shape-based and color-based features along with learning-based methods (Convolutional Neural Networks (CNN)). Subsequently, the signs were recognized using selected techniques such as Random forest method, CNN, and Support Vector Machine (SVM). Selected techniques were applied to the 10 varieties of traffic signs in Sri Lanka in different conditions, each having 1000 samples. Experimental results show that the approach obtained the desired results effectively. CNN method obtained 74.16% overall accuracy, SVM method obtained 63.5% overall accuracy and Random forest method obtained 58.6% overall accuracy. In the future, accuracy can be improved by testing the technologies in different internal factors like different camera quality (fps) and different computing power, as well as high-resolution images and a large number of training images should be used for the analysis. The experimental results showed that CNN is the most suitable technology to detect and recognize traffic signs based on the Sri Lankan traffic signs database Keywords: Traffic sign detection and recognition; Convolutional Neural Network; Support Vector Machine; Shape based methods; Color based methods; Random forest
  • No Thumbnail Available
    Item
    Detection and Classification of Diseased Tomato Leaf Using Image Processing Techniques
    (Uva Wellassa University of Sri Lanka, 2016) Mehendran, Y.; Kartheeswaran, T.; Ediriweera, E.P.S.K.
    Tomato plants are highly vulnerable to fusarium wilt, verticillium wilt, and late blight. The symptom is yellowing of the lower leaves, which gradually wilt and die. The naked eye observation of experts is the main approach adopted for detection and the identification of plant diseases. We developed a method to detect and classify damages in leaves using image processing techniques. For this experiment, images downloaded from the interne were used. The disease regions were segmented using K-Means clustering and the classification of the disease was done with Support Vector Machine (SVM) by training with the selected features from the training set of images. The initial version has three classes such as Bacterial Wilt, Early blight and Healthy tomato leaves. The accuracy level for the identification and the classification of diseases was calculated for each category separately. The accuracy of the system for the selected nine features was calculated as 76.5%. Association among the features as Contrast, Correlation, Energy, Homogeneity, Entropy, Mean, Standard deviation, Skew, and Kurtosis gave the optimum accuracy. This system with high accuracy motivates the other researchers to extend the system with added functionality, which will be a farmer friendly software solution. Keywords: HSI, K-means, Gray-level co-occurrence matrix, Support Vector Machine (SVM),
  • Thumbnail Image
    Item
    Distinguish Garnet Mineral from Pulmuddai Beach Sand Using Image Processing Techniques
    (Uva Wellassa University of Sri Lanka, 2019-02) Hirosh, D.M.H.; Darshana, J.V.A.; Doratiyawa, H.M.M.S.; Wilson, R.S.I.; Mehendran, Y.; Jaliya, R.G.C.
    Beach sand is one of the major minerals producing source in Sri Lanka. Pulmuddai Beach, rich in Ilmenite, Rutile and Zircon and Garnet and it is the largest mineral processing plant in Sri Lanka. It's a great necessity to explore high mineral localities for production. In the industrial level, the percentage of a specific mineral is calculated using a visual inspection through a microscope which is manual and time-consuming. The research introduces an innovative method to distinguish Garnet mineral from sand using image processing techniques. In this study, 1125 visible light RGB (Red, Green, Blue) images and 1125 Infrared (IR) images of beach sand were captured in a controlled light environment. RGB color composite images and IR images were analyzed separately to identify Garnet mineral from the gang and to calculate Garnet percentage. For the machine learning classification purpose, contrast, variance, mean, median, min, max, range, kurtosis, skewness, standard deviation and correlation were extracted from sand grains images. Then RGB, HSV (Hue, Saturation, Value) and RGBIR (Red, Green, Blue and Infrared) color models were used through a machine learning model. The highest accuracy of 63% of separation accuracy was given by the HSV color model. The accuracy could be increased by introducing more images to the machine learning process. The final model was built based on the HSV color model since it has the high accuracy of separation. Then the HSV model subjected to object counting model, area-based counting model and volume based counting model to identify the most suitable method for the percentage calculations. Among these three methods, an object counting model produced the more accuracy results with 57%. Thus, the HSV color model incorporates with object counting model produces the best combination to identify Garnet and calculate its percentage.
  • Thumbnail Image
    Item
    Mobile Apps’ Feature Extraction Based On User Reviews Using Machine Learning
    (Uva Wellassa University of Sri Lanka, 2019-02) Thiviya, T.; Nitheesram, R.; Srinath, G.; Ekanayake, E.M.U.W.J.B.; Mehendran, Y.
    The star rating and user reviews of Google Play store play a major role in App Store Optimization. The average number of stars received for an app and the user reviews are used to evaluate the overall app quality. We argue that the star rating is not a reliable measurement for the user satisfaction, since the star rating is a straightforward mathematical expression only. We cannot find a user’s real experience about the app by asking them to rate the app out of five stars, even rating user reviews will give only an overall user perspective about the app. Therefore, we recommend a specific app feature evaluation method based on user reviews, which give us a genuine app rating than the conventional method. In addition to review based rating we did popular feature extraction from user reviews. Initially we started our research with mining user reviews from Google Play Store for several categories by using Web Scraping tool. We used the sentiment analysis to extract the meaning from the reviews and define the polarity of them. According to the polarity strength we rated each reviews. Overall rating was calculated by finding the average of given review based ratings. Then we compared review based ratings with the existing star ratings. We found that compare to overall review ratings, the existing star ratings differs and high. Moreover, the app feature set was selected according to the category of the app, and then the popularity of those features was calculated by using machine learning. We were resulted with app features and their popularity based on users' reviews. These popularities can be helpful to the potential app users to get know the top features and their popularity of the particular apps. In the meantime, app developers can identify which features have low popularity and they can improve those features in future.