Image splicing¶
Image splicing is the process of combining regions from multiple source images to create a composite image. A splicer typically cuts out a region from one image and inserts it into another, followed by post-processing (blending, color correction, boundary smoothing) to make the result appear seamless. Splicing is common in satire, misinformation, and manipulated media.
Variants¶
Face splicing: Replacing a person's face with another face from a different source image. Often includes resizing, rotation, and illumination adjustment to match the target context.
Regional splicing: Inserting arbitrary image regions (objects, backgrounds) rather than faces.
Copy-move forgery: Copying and pasting a region from one location to another within the same image (e.g., duplicating a person, removing an object).
Detection methods¶
Boundary analysis: Splicing boundaries often exhibit sharp discontinuities in color, texture, or lighting that can be detected via edge detection or gradient analysis.
Noise inconsistency: Different source images have different in-camera noise characteristics (due to CFA patterns, ISO settings, sensor properties). When spliced, the tampered region's noise pattern differs from the background. Techniques include local noise estimation and CFA pattern analysis.
Double JPEG compression: If the original image was JPEG-compressed before splicing, the background will show double JPEG artifacts while the newly spliced region (typically inserted and re-saved) may have different compression patterns.
Luminance/illumination: Spliced regions may show different light source directions or global illumination properties than the background.
Deep learning: CNN-based methods trained on splicing datasets (e.g., CASIA, Columbia) can learn to classify images as authentic or forged. These often operate on ELA (Error Level Analysis) preprocessed images.
Key papers¶
- Two-Stream Neural Networks for Tampered Face Detection — specifically targets face splicing via two-stream network combining visual artifact detection with steganalysis features.
- A Multi-Modal Method for Satire Detection using Textual and Visual Cues — applies image forensics to detect manipulated images in satirical news.
Datasets¶
- CASIA 1.0 and CASIA 2.0: Splicing and copy-move forgery datasets
- Columbia Image Splicing Detection dataset: 180 spliced and 180 authentic images
- SwapMe and FaceSwap Dataset — face splicing via face swapping
Limitations and challenges¶
Post-processing: High-quality splicing includes seamless blending, boundary reconstruction, and color matching that can defeat forensic techniques based on boundary discontinuities or simple compression artifacts.
Recompression: Images downloaded and re-uploaded on social media undergo recompression, which can obscure JPEG artifacts and noise pattern inconsistencies.
Limited dataset diversity: Most splicing datasets are relatively small and may not capture the diversity of splicing techniques, image content, and quality levels found in real-world misinformation.
Connection to misinformation¶
Image splicing is a key technique for creating visual misinformation. By combining faces or objects from different contexts, splicing can create false claims of association, events, or statements. For example, a politician's face can be spliced into a compromising context to create misleading political content.
Connections¶
- Image forensics — general tampering detection framework
- Face tampering detection — specialized to face regions
- Visual misinformation — using manipulated images in misinformation campaigns