Automatic image forgery classification using Deep Learning technique with Block matching process

Abstract

This study addresses the difficult task of automatically classifying fake images by using state-of-the-art deep learning (DL) methods in conjunction with a thorough block-matching procedure. Robust approaches are required to distinguish genuine information from fakes due to the increasing prevalence of picture alteration. In response, a multi-stage strategy is being developed that focuses on combining DL approaches for feature extraction and classification. These methods include ResNet- 50, VGG16, and Convolutional Neural Networks (CNN), all of which include pre- trained weights. On the CASIA V2 dataset, ResNet-50 achieves an impressive 93.5% accuracy, which is worth noting. Given the intricacy of the task, more conventional methods like Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN) are also utilized. The RF model stands out with an 88% accuracy in photo authentication, whereas the other approaches display various recall, accuracy, and F1-scores. The Ensemble Classifier improves prediction accuracy even further, and it successfully identifies real photos most of the time. With 99% accuracy on the MICC-F220 dataset and 98% accuracy on the CASIA V2 dataset, the study closes by presenting a lightweight DL approach employing the U-Net model with pre-trained weights. The technique demonstrates remarkable performance. Finally, by shedding light on the relative merits of DL and more conventional approaches, this study considerably progresses the state of the art in automatic picture forgery classification. While laying the framework for further research and advancements in the field of automatic image classification, the suggested methods also help to improve photo counterfeit detection. newline