Image Copy-Move Detection Based on Hierarchical Enhancement and Cross-Model Correlation Representation Fusion

Abstract

This paper tackles the pervasive issue of image copy-move forgery by introducing a novel method, Hierarchical Enhancement and Cross-Model Correlation Representation Fusion (HECF-Net), which leverages multi-model and multi-level coordination to accurately identify copied and tampered regions within images. To mitigate shallow information loss in deep neural networks (DNNs) and reduce false positives caused by low edge texture differentiation in original images, we develope a Hierarchical Edge Enhancement (HEE) module. The module introduces a novel Cross-Dimensional Hierarchical Enhancement approach, which employs the Scharr edge detection algorithm for edge computation and achieves cross-dimensional fusion within both single-channel and multi-channel layers of Convolutional Neural Networks (CNNs). Additionally, we propose a filtering-based Swin-Transformer (FST) architecture to extract global features from real images. By utilizing a channel attention mechanism, the model selectively emphasizes similar information within the broader feature set. The Cross-Model Correlation Fusion (CMCF) module integrates multi-scale and multi-level information from local, edge, and global features, performing Pearson correlation calculations and fusion to produce robust similarity feature maps. Extensive experiments demonstrate that our method outperforms most current advanced techniques on publicly available benchmark datasets, confirming its significant advantages.

Keywords: Image Manipulation, Copy-Move Detection, Convolutional Neural Networks, Swin-Transformer, Pearson Correlation