Abstract
Image forgery is the practice of tampering with or changing an image. It could seem safe to manipulate images in this way. However, to preserve the credibility of the evidence, images used in court or forensic investigations must remain unmodified. Forgery detection can be carried out utilizing a variety of different ways, just like picture modification. For this goal, two distinct algorithms are proposed in this paper, and the superior approach is chosen. Choosing the most suitable deep-learning algorithm to utilize in the process of spotting image counterfeiting is the study's major goal. The CASIA Tampered Image Detection Evaluation dataset includes both unedited and modified photos. Denoising, scaling, and resizing are used in the initial preprocessing of the photos. The preprocessed dataset is then used to train the deep-learning models. Deep learning (DL) models are built using the UNet and ResNet-UNet, two separate methods. Both models showed similar outcomes during training. But even from the first epoch, the ResNet-UNet method produced better results. The models are examined to find the optimum algorithm after seeing the same results during training. This testing is based on a wide range of factors, including but not restricted to precision and accuracy. The model created using the UNet algorithm has an acceptable final accuracy of 94.6. But the ResNet-UNet method has a 96.3% accuracy rating, which is higher. The ResNet-UNet algorithm produces better outcomes in terms of all the other parameters, similar to how accurate it is. As a result, the ResNet-UNet model is determined to be the best method and is implemented into a software application's backend process where it successfully forecasted a fake image.
Keyword
Image Manipulation, Deep Learning, Image Preprocessing, Forgery Detection, Accuracy, Mobile App
PDF Download (click here)
|