Effective Deep Features for Image Splicing Detection

Abstract

In the last few years, Digital image forgery (DIF) detection has become a prominent subject. Image splicing is a frequent approach for making digital image forgeries. Image splicing creates forged images that are hard to detect immediately. The detection accuracy of most existing image splicing detection algorithms is low, thus there is room for improvement. Therefore, this research provides an image splicing detection (ISD) method based on deep learning. The proposed image splicing detection has three stages: (1) RGB image conversion and image size fitting are examples of image pre-processing. (2) Using the pre-trained CNN AlexNet model, we extract the final discriminative feature for a preprocessed image. (3) Finally, the generated feature representation is used to train a Canonical Correlation Analysis (CCA) classifier for binary classification (authentic/forged). The accuracy of the proposed approach using a pre-trained AlexNet model based deep features with CCA classifier is equal to 98.79 % when evaluated on the CASIA v1.0 splicing image forgery database. In comparison, the proposed surpassed existing methods. In the future, the proposed could be applied to other types of image forgery, such as image retouching.