A Multidomain Deep Learning Approach for Accurate Detection and Localization of Copy-Move Forgeries using CLMD-BFA

Abstract

The problem of image forgery is defined in terms of detecting, classifying, and localizing manipulations. Existing forgery localization techniques have been found to be limited either in terms of particular types of forgery they can handle or the need for increased computational cost when dealing with multiple manipulations. In order to solve such problems, this paper presents an efficient cascaded learning-based approach to multimodal image forgery localization through differential bio-inspired feature analysis (CLMD-BFA). In the proposed approach, the initial segmentation of forged/real image regions is carried out using the U-Net architecture and subsequent classification of forgery types is done through cascaded learning. The multi-modal features like frequency, entropy, Gabor, wavelet and convolutional features are obtained and optimized through Grey Wolf Optimizer (GWO). As compared to the latest techniques in image forgery detection and localization such as CW-HPF, DCNN, MSTA-Net, and transformer-based architectures, the proposed CLMD-BFA framework offers a more advanced multimodal feature representation and increased localization ability while keeping computational efficiency intact for varied forgery types. The optimized common spatial and temporal patterns (CSTPs) are used to train a 1D-CNN for patch-level forgery localization. Experimental results on several datasets with multimodal image forgery show the effectiveness of the proposed technique to detect and localize forgery accurately. Through comparative analysis, it is observed that the proposed approach performs better than traditional hand-crafted feature based techniques and modern deep learning based approaches.