Image Forgery Detection using Image Denoising and Cascaded 3D UNET Segmentation

Abstract

Image forgery detection is the identification of changes or modifications in digital images, thereby guaranteeing their integrity and validity. It is very important in the fight against false information, online fraud, and maintaining confidence in visual media. Multi-Noise Elimination Neural Network (MNEN) is introduced as a novel multi-scale nested encoder network. The goal is to reduce noise while maintaining appropriate structural integrity. MNEN successfully captures spatial hierarchies and contextual interactions by using a multi-scale, layered architecture that collects data at several levels. In the encoder pipeline, highly connected convolutional layers with skip connections improve suspicions about fading gradients while still providing considerable feature propagation. Using a layered skip fusion approach to merge low- and high-level semantic data helps to minimize noise while maintaining image quality. Blocks of progressive upsampling and filtering allow the decoder pipeline to recover high-quality denoised outputs. Extensive testing on benchmark datasets reveals that, in terms of PSNR and SSIM, MNEN frequently exceeds state-of-the-art techniques. Results of ablation studies reveal that denoising performance is much improved by layered connections and multi-scale feature integration. Because of its competitive inference times, MNEN is ideal for implementation on edge devices and real-time applications. Studies indicate that MNEN is a potential technique for image restoration in various contexts. The Cascaded 3D U-Net model enhances digital image tampering detection by performing multi-level segmentation on image data