Open Access

ARTICLE

Robust Image Forgery Localization Using Hybrid CNN-Transformer Synergy Based Framework

Sachin Sharma^1,2,*, Brajesh Kumar Singh³, Hitendra Garg²

1 Department of Computer Science, Abdul Kalam Technical University, Lucknow, 226031, India
2 Department of Computer Engineering & Applications, GLA University, Mathura, 281406, India
3 Department of Computer Science, Raja Balwant Singh Engineering Technical Campus, Agra, 283105, India

* Corresponding Author: Sachin Sharma. Email:

Computers, Materials & Continua 2025, 82(3), 4691-4708. https://doi.org/10.32604/cmc.2025.061252

Received 20 November 2024; Accepted 21 January 2025; Issue published 06 March 2025

Abstract

Image tampering detection and localization have emerged as a critical domain in combating the pervasive issue of image manipulation due to the advancement of the large-scale availability of sophisticated image editing tools. The manual forgery localization is often reliant on forensic expertise. In recent times, machine learning (ML) and deep learning (DL) have shown promising results in automating image forgery localization. However, the ML-based method relies on hand-crafted features. Conversely, the DL method automatically extracts shallow spatial features to enhance the accuracy. However, DL-based methods lack the global co-relation of the features due to this performance degradation noticed in several applications. In the proposed study, we designed FLTNet (forgery localization transformer network) with a CNN (convolution neural network) encoder and transformer-based attention. The encoder extracts local high-dimensional features, and the transformer provides the global co-relation of the features. In the decoder, we have exclusively utilized a CNN to upsample the features that generate tampered mask images. Moreover, we evaluated visual and quantitative performance on three standard datasets and comparison with six state-of-the-art methods. The IoU values of the proposed method on CASIA V1, CASIA V2, and CoMoFoD datasets are 0.77, 0.82, and 0.84, respectively. In addition, the F1-scores of these three datasets are 0.80, 0.84, and 0.86, respectively. Furthermore, the visual results of the proposed method are clean and contain rich information, which can be used for real-time forgery detection. The code used in the study can be accessed through URL: (accessed on 21 January 2025).

---

Keywords

---