Using Patch Analysis on Detecting Seam Carved Images

Seam carving is a content-aware digital image resizing algorithm. Pixels on a digital image are assigned with energy for later connection into seams. Repeatedly removing seams that are with minimal energy, we can reduce the width (or height) of the image. This way seam carving retains important visual information while resizing an image implies that we can assign lowest energy value to a specific region on the image and thus eliminate particular objects deliberately. Development of a seam carving detection method is extremely important, although challenging. Herein, we propose a novel method, referred to as "patch analysis", for detecting seam-carved images. Images are divided into super pixels and then searched for one of nine types of patches that can possibly recover a super pixel from seam carving. Our method analyzes the patch transition probability among three-connected (sub-diagonal, vertical and main diagonal) super pixels. Transition probability provides Markov features for a Support Vector Machine (SVM) to construct a classification model. Our rough experiment indicates that the proposed method is potentially the best seam carving detection method in current stage. Our plan in this two-year project is as follows: In the first year, we will complete all the necessary experiments and thus to examine and refine our model. In the next year, we will extend our method to identify the hot regions frequently crossed by carved seams.

Project ID:PB10108-2816
External Project ID:NSC101-2221-E182-073