A Comparative Analysis of LBP Variants for Image Tamper Detection

This thesis explores the use of Local Binary Patterns (LBP) and Convolutional Neural Networks (CNN) for detecting image tampering, an increasingly prevalent issue in today's digital landscape. Through a comparative analysis of four LBP variants using the CASIA-2.0 dataset, it combines LBP's texture descriptors with CNN to enhance accuracy and robustness. The methodology involves generating local texture descriptors with LBP and feeding them into a CNN architecture trained to classify images as tampered or authentic. Despite challenges like computational complexity, the research aims to contribute to a reliable tamper detection system applicable in various real-world scenarios. Notably, Uniform LBP demonstrates superior performance in both training/testing time, achieving accuracy and F1-score exceeding 97% in image tamper detection, validating the effectiveness of the approach.