A Deep Learning Framework with Transfer Learning Approach for Image Forgery Localization

Altering digital content is one of the major concerns experienced in the digital environment. Because of the developments in image modification applications in this digital era, the tampering of images for harmful reasons has proliferated, necessitating the detection of such manipulation before it misleads in any of the applications. A framework based on DeepLabv3+ is presented in this paper for detecting several forms of forgeries and locating the altered areas precisely within the images. To pre-process the forged photos, Error Level Analysis (ELA) is used, which differentiates the areas with varying quantization error levels. The complete framework is trained using the pre-processed images as input. The encoder side of the DeepLabv3+ architecture uses a pre-trained ResN et50 network for transfer learning. DeepLabv3+ uses Atrous convolution that helps to get the broader field of view and extracts contextual features for enhancing the precision of localization. At the decoder, the output features of the encoder are merged with the low-level features of the input image to appropriately locate the forged area. With high precision, this framework can identify and locate many types of manipulations. The findings illustrate that the proposed framework outperforms several existing forged area localization techniques.