Using Deep Network Mapping and Labeling Techniques to Locate the Modified Regions in a Seam Carved Image

Seam carving is a unique image processing method, which has been successfully used in not only resizing an image but also removing particular objects from the target image. Based on content-aware energy computing, seam carving can cause hard-to-recognize tampering of images. There have been various methods proposed, with increasing accuracy rates, to detect seam carved images; meanwhile, identifying the modified regions is regarded as an advanced and even more challenging research direction. Although difficult, this research topic is important due to its potential to help forensic personnel in judging the intention of image modification, presuming the original content before the change, and persuading the suspects into confession and cooperation. In past work, we have ever tried deconvolution neural networks to locate the suspicious seam-carved area. Experimental results revealed that using deep learning methods could be feasible to this research direction. We then surveyed other two novel techniques of deep learning, i.e., Class Activation Mapping (CAM) and Region of Interesting (ROI) Labeling, and found that they are significantly promising for searching the evidences of seam carving. We therefore propose this two-year project, aiming to use the CAM mapping and ROI labeling methods, respectively, to identify the seam carved regions in a modified image. During the research period, particular deep neural networks will be developed to search for and compare the characteristics of the image contents. The research achievements will be interesting and potential for high-quality academic publication. Deep learning techniques are now widely used in labeling objects with specific shapes. Nevertheless, the proposed research project is going to label something that does not exist. This project is thus not only practical for digital forensics but also attractive in academic discussion.

Project ID:PB10907-4769
External Project ID:MOST109-2221-E182-051