DEVELOPMENT OF GAUSSIAN COPULA CLASSIFICATION MODEL WITH SYMBOLIC DATA DISTRIBUTION FUNCTION APPROACH (CASE STUDY: IMAGE CLASSIFICATION)

• Main Author: Winarni, Sri
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79529
• Institution: Institut Teknologi Bandung

This dissertation proposes an innovative approach in image classification using Gaussian Copula with symbolic data representation to model images as distributions of pixel intensity and location values. The approach aims to address limitations commonly found in image classification methods that rely on pixel intensity features, leading to high memory requirements and computational complexity. By modeling with symbolic data approach in the distribution function, these issues can be effectively reduced. This dissertation suggests a simple and efficient image classification model by modeling the dependency between random variables in the image using Gaussian Copula, along with feature extraction from the distribution of pixel intensity and location. The research methodology includes image preprocessing, feature selection, Gaussian Copula modeling, and classification model training. This dissertation also proposes a new test statistic, namely the chi-square test statistic based on Gaussian Copula, for image classification. The approach is implemented on the MNIST database, and the results are measured using accuracy. The research outcomes include the development of a simpler and more accurate image classification model, abstract feature extraction, and a new approach to determining image classes using test statistics. The research findings reveal that the distribution of location plays a crucial role in describing the characteristics of image features, alongside the distribution of pixel intensity. Moreover, the number of partitions also influences the accuracy of the classification obtained. Four-row partitions yield better accuracy results compared to three and two-row partitions, as well as full data. The precise differentiating points also contribute to determining the performance of the classification model. The best Gaussian Copula classification model is obtained with pixel intensity and location features in four-row partitions and two differentiating points, achieving high accuracy .. This research has the potential to make a significant contribution to the development of improved and efficient image classification methods.