CCLHF-Net: Constrained Convolution Layer and Hybrid Features-Based Skip Connection Network for Image Forgery Detection

Nowadays, images are playing a crucial role in various important applications like medical, forensic, and legal. However, if someone tampers with these images, they become invalid. Unfortunately, the increasing popularity and ease of image editing tools have made it easier to manipulate images. Therefore, it is crucial to be able to detect any such image manipulations to maintain the integrity of the images. Several techniques have been presented to detect these image forgeries in the literature. This paper proposes a deep learning architecture that uses an encoder–decoder network with a long short-term memory to localize the tampering operations of an image. The encoder network comprises a constrained convolution layer followed by a fully convolutional network with ResNet-50 as its backbone to extract spatial domain feature maps. The long short-term memory network that learns the transition between altered and non-altered image region feature maps in the frequency domain is exploited. The feature maps from long short-term memory and encoder are fused, and these low-resolution feature maps are given as input to the decoder network, which learns to map pixel-wise predictions of an image tamper localization. Skip connections are added from the encoder’s convolution network to the decoder’s corresponding level to achieve better semantic consistency across layers. The proposed architecture is experimented on two benchmark datasets, DEFACTO and CASIAv1, and achieves an area under curve score of 0.66 and 0.84, respectively. These results show that the proposed architecture could perform better than the state-of-the-art image forgery models.