Exposing image splicing traces in scientific publications via uncertainty-guided refinement

Area of Science:

• Computer Science
• Image Forensics
• Scientific Publishing

Background:

• A rise in manipulated scientific images has led to retractions, emphasizing the need for image integrity.
• Existing forensic tools address image duplication and synthesis but not splicing, which is more complex.
• Detecting spliced images in scientific contexts is hindered by artifacts, noise, and a lack of specialized datasets.

Purpose of the Study:

• To address the challenge of detecting image splicing in scientific publications.
• To mitigate disruptive factors in scientific images that complicate splicing detection.
• To introduce a novel method and dataset for advancing image splicing detection research.

Main Methods:

• Development of an uncertainty-guided refinement network (URN) designed to handle image artifacts and noise.
• Construction of a new dataset, SciSp, comprising 1,290 manually spliced scientific images.
• Rigorous experimental evaluation of the URN's performance on the SciSp dataset.

Main Results:

• The proposed uncertainty-guided refinement network (URN) demonstrated superior performance in detecting image splicing.
• The SciSp dataset provides a valuable resource for training and evaluating image splicing detection models.
• The URN effectively mitigates disruptive factors present in scientific images.

Conclusions:

• The URN offers a promising solution for enhancing the integrity of scientific publications.
• Further research in image splicing detection is essential to combat scientific misconduct.
• The developed dataset and method contribute significantly to the field of image forensics.