DSNI-Net: Forensic Network for Detecting Splicing in Salt-and-Pepper Noisy Images

Image splicing forensics technology is an important way to detect and locate spliced areas in naturally tampered images. However, common noise, especially salt-and-pepper (s&p) noise, can change pixel values to 0 or 255, significantly reducing the accuracy of splicing forensic methods. As the noise intensity increases, the impact on forensic methods gradually increases. This paper proposes a forensic method to enhance the robustness of the forensic method in scenarios with high-intensity s&p noise called forensic network for detecting splicing in s&p noisy images (DSNI-Net). The image is first denoised using a combination of residual network and median layer (ML) before forensics. The CNN network extracts shallow and deep feature channels in the image, uses ML to remove s&p noise, and reconstructs the image content through the convolution structure to restore image details. Secondly, the feature comparison structure is used to extract tampering traces during the forensic process. This ensures that the final extracted features are uniquely related to the image itself and are used to reveal areas of image tampering. Experimental results on noisy datasets demonstrate the robustness and effectiveness of DSNI-Net under various intensities of s&p noise. Notably, at a noise intensity level of 5%, DSNI-Net improves forensic performance by 37%. Generalization experiments further confirm that DSNI-Net is robust against other types of noise and JPEG compression.