Attention-Based Dual-Domain Fusion Network for Median Filtering Forensics
Yakun Niu, Lei Tan, Lei Zhang, Yingjian Chen, Xianyu Zuo · IEEE Signal Processing Letters · 2024
Median filtering forensics has attracted much attention in recent years. However, most existing methods exploit either spatial domain or frequency domain features, while neglect the correlation between them. Moreover, they often suffer from inadequate generalization in cross-database and cross-parameter. To solve these problems, we propose an attention-based dual-domain fusion network (DDFNet) to fuse spatial domain and frequency domain features in a mutually complementary way. In the spatial domain, we use four types of modules aimed at extracting the pixel residual features. In the frequency domain, the discrete cosine transform (DCT) coefficients are fed into a residual network to capture the frequency residual features. Finally, the features obtained from the two domains are fused by an attention module to mine their latent complementary relationships. Extensive experiments demonstrate that the proposed DDFNet outperforms the state-of-the-art methods in terms of robustness and generalization capacity.