Implementation of the Learning Vector Quantization (LVQ) Method with GLCM Extraction for Image-Based Fake News Detection with Text Overlay

Majid, Vito Fausta (2026) Implementation of the Learning Vector Quantization (LVQ) Method with GLCM Extraction for Image-Based Fake News Detection with Text Overlay. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) 20081010109.cover.pdf Download (467kB)
	Text (Bab 1) 20081010109.-bab1.pdf Download (357kB)
	Text (Bab 2) 20081010109.-bab2.pdf Restricted to Repository staff only until 17 July 2029. Download (497kB)
	Text (Bab 3) 20081010109.-bab3.pdf Restricted to Repository staff only until 17 July 2029. Download (832kB)
	Text (Bab 4) 20081010109.-bab4.pdf Restricted to Repository staff only until 17 July 2029. Download (589kB)
	Text (Bab 5) 20081010109.-bab5.pdf Download (336kB)
	Text (Daftar Pustaka) 20081010109.-daftarpustaka.pdf Download (337kB)

Abstract

Fake news (hoax) based on images with overlaid text is one of the most rapidly spreading forms of disinformation on social media, with significant negative impacts on public perception. Automatic detection of this type of manipulation is particularly challenging, as the textural changes introduced by text overlay are subtle and difficult to distinguish visually. This study implements the Learning Vector Quantization (LVQ) method combined with Gray Level Co-occurrence Matrix (GLCM) feature extraction to detect image-based fake news. Images are processed through a preprocessing stage involving grayscale conversion and normalization, after which texture features are extracted using GLCM across four angular directions (0°, 45°, 90°, 135°), yielding four primary features contrast, energy, correlation, and homogeneity resulting in a 16-dimensional feature vector. The LVQ model was built from scratch and trained to classify images into two classes: Original and Overlaid. Testing was conducted through four parameter variation scenarios covering the number of prototypes per class, initial learning rate, maximum number of epochs, and decay rate. The best model was obtained with parameters n_prototypes=10, learning_rate=0.1, max_epochs=100, and decay_rate=0.99, achieving an accuracy of 80.50%, precision of 81.02%, recall of 80.50%, and F1-Score of 80.42% on the test data. The model was subsequently deployed using the Streamlit framework as an interactive web application, enabling users to upload images and obtain classification results automatically. The results demonstrate that the combination of LVQ and GLCM is capable of adequately detecting image-based fake news with overlaid text, and can be further improved through the addition of color features or more advanced LVQ variants.

Item Type:

Thesis (Undergraduate)

Contributors: