DeepCMFD

Abstract

In this paper, we propose a new method to detect and localize a copy-move forgery using feature descriptors based on a deep neural network. The success of deep neural networks in other computer vision works operates in a way that understands a semantic meaning of a region of interest only in an image. In contrast, the proposed network additionally capture detailed characteristics of the whole image including the backgrounds. We modify the last layer to fully convolutional, allowing generation of feature per pixel unit. Furthermore, fast GPU-based kNN search introduced in a matching step is as fast as the keypoint-based counterparts, while ensuring a remarkable performance of the dense fields. For post-processing procedure, we implemented DBSCAN and morphological technology to remove false positives. The experiments using a benchmark dataset show that our method is more robust against JPEG compression, additive Gaussian noise, rotation, and scale within a certain range than comparison techniques.