A patch analysis method to detect seam carved images

• We propose a patch analysis approach to detecting seam-carved images.
• Images are divided into 2 × 2 blocks and the optimal patch type for each block is selected.
• Analyzing the patch transition probability achieves currently best detection accuracy.
• This method can be extended to identify the hot regions frequently crossed by carved seams.

Seam carving is a content-aware image processing algorithm that has been successfully applied to resizing and deliberately removing objects from digital images. Retargeting images by seam carving is hard to identify; therefore, the detection of seam-carved images has been an important and attractive research topic. Existing methods for detecting seam-carved images include those derived from steganography attacks and those based on statistical features. However, these algorithms leave scope for further improvement. Here, we propose a novel method in which images are divided into 2×22×2 blocks, referred to as mini-squares  , and then searched for one of nine types of patches that is likely to recover a mini-square from seam carving. Our method analyzes the patch transition probability among three-connected mini-squares and achieves currently best detection accuracies, namely, 92.2%92.2% and 95.8%95.8% for 20%20% and 50%50% seam-carved images respectively. We also discuss in this paper other potential applications of our patch analysis method, for example, identification of the hot regions frequently crossed by carved seams.