Smart Computing and Systems Engineering - 2025 (SCSE 2025)

Permanent URI for this collectionhttp://repository.kln.ac.lk/handle/123456789/30037

Browse

Author: search.filters.author.Kumara, B. T. G. S.Has files: Yes

Search Results

Now showing 1 - 3 of 3
  • No Thumbnail Available
    Item
    Deep Learning-Based Approach for Distinguishing Between AI-Generated and Human-Drawn Paintings
    (Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Warnakulasooriya, A. I.; Rupasingha, R. A. H. M.; Kumara, B. T. G. S.
    With the increasing number of robust Artificial Intelligence (AI) art generation applications, more realistic AI-generated paintings (AIGPs) are emerging, creating a significant impact on artists. Due to the widespread acceptance of AIGPs, the cultural, historical, and monetary value of real masterpieces is becoming uncertain, raising concerns about the significance of human painters and their artistic techniques. To protect artists’ rights, it is crucial to differentiate AIGPs from human-drawn paintings (HDPs). Accordingly, the main objective of this research is to develop a Convolutional Neural Network (CNN) model that can automatically distinguish between AI-generated and human-drawn paintings without human intervention. Unlike previous studies that focused mainly on pixel-level analysis, the proposed model considers additional features such as edge patterns, object arrangements, pattern distributions, and gradient characteristics in painting classification. A diverse dataset of 3,000 paintings from the AI-ArtBench Dataset—comprising 1,500 AIGPs and 1,500 HDPs across 10 different art themes—was collected and preprocessed for this study. The AIGPs were generated in equal proportions using Latent Diffusion and Standard Diffusion Models. The implemented CNN model achieved an optimum classification accuracy of 90% with a training data size of 10%, while the ANN model exhibited 77% accuracy under the same conditions. Furthermore, models were compared using performance metrics such as precision, recall, F1-score, RMSE, and MAE. Through Gradient-weighted Class Activation Mapping (Grad-CAM), the key visual features that the CNN model used to distinguish AIGPs from HDPs were identified. These findings highlight the potential of automated systems in detecting AI-generated versus human-created artworks for authentication purposes. Future work will focus on analyzing model performance across different art styles and identifying the unique discriminative features associated with each.
  • No Thumbnail Available
    Item
    AI-Driven Solutions for Automated Fish Freshness Classification Using CNN Architectures
    (Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Peries, R. F. S.; Adeeba, S.; Ahamed, M. F. S.; Kumara, B. T. G. S.
    Ensuring fish freshness is essential for market value, consumer health, and seafood quality. In Sri Lanka, traditional sensory-based methods for assessing freshness are subjective and often inaccessible to small-scale fishermen due to high costs and limited resources. This study addresses these challenges by employing Convolutional Neural Networks (CNNs) to automate fish freshness classification using image data from the Mannar coastal region. The approach involved capturing images of whole fish, fish eyes, and fish gills, followed by preprocessing steps such as labeling, resizing, and augmentation. Separate custom CNN models were developed for each dataset, with the gill dataset achieving the highest performance at 98.26% accuracy, along with excellent precision, recall, and F1-scores. Furthermore, advanced pre-trained models—including VGG16, ResNet50, MobileNetV2, InceptionV3, Xception, and DenseNet121—were evaluated on the gill dataset. Among these, DenseNet121 emerged as the best-performing model due to its high accuracy, precision, recall, F1-score, and stable learning curve. These findings highlight the potential of CNN-based and pre-trained models to provide scalable, cost-effective solutions for fish freshness assessment, promoting sustainable seafood practices, empowering small-scale fishers, and enhancing food safety standards.
  • No Thumbnail Available
    Item
    A Deep Learning-Based Approach for Detecting Duplicate GitHub Issues in Open-Source Repositories Using LSTM
    (Department of Industrial Management, Faculty of Science, University of Kelaniya., 2025) Dharmadasa, T. K. R. S.; Rupasingha, R. A. H. M.; Kumara, B. T. G. S.
    GitHub is a platform used along with the popular version control tool Git to provide hosting facilities to software repositories. Users can publish GitHub issues to notify the repository contributors about bugs, questions, and feature requests. GitHub hosts open-source repositories that are contributed by developers across the globe. The asynchronous and uncoordinated nature of these contributions in open-source repositories increases the probability of posting duplicate GitHub issues, resulting in redundant efforts. The standard mechanism introduced by GitHub to mark duplicate issues is adding a comment to that issue body mentioning the original issue. Then GitHub will add the corresponding duplicate tag and close that issue. However, due to manual labor required to find duplicates, developers are discouraged from seeking similar issues before publishing a new issue to GitHub. The study’s main objective is to address this problem and propose an automated solution using deep learning algorithms. Our research introduces a novel approach that combines feature extraction and similarity calculations to identify duplicate GitHub issues. The proposed methodology extracted over 4000 GitHub issues covering different programming languages and repositories. After pre-processing, various features were extracted using multiple feature extraction techniques, and semantic similarity metrics such as cosine similarity were utilized to create the feature vector. The feature vector was used with different algorithms like Artificial Neural Network (ANN) and Recurrent Neural Network (RNN) including deep-learning algorithms like Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN). Algorithm results are compared to detect the most suitable approach for detecting duplicate GitHub issues. Based on the different evaluations, LSTM is the better approach resulting in 88% accuracy with the highest precision, recall, and f-measures while giving the lowest error rates. With this proposed methodology, duplicate GitHub issues can be easily detected, reducing the manual work.