Detection of seam carving based size modification in digital images

Seam carving is one of the mostly applied content-aware image resizing methods today. The deteriorations caused by seam carving are mostly unnoticeable for human eyes unless the scaling ratio is very high. The reason of this visual success comes from evaluating the pixels according to their importance values. Optimal seam of an image is a pixel path which contains the least energy (importance) throughout the image. Image width or height is decreased by one in each iteration by removing those unimportant, one-pixel width pixel paths. The semantically important foreground objects remain untouched as far as possible. Since such a scaling approach which perfectly preserves the image content can be used malevolently, the detection of the images that are scaled in this manner becomes more of an issue. The detection of seam carving is more difficult than the other scaling methods since evaluating the images geometrically is not sufficient, but a detailed analysis investigating the semantical concept is required. In this study, the detection of the images scaled by seam carving is realized by feature extraction and training a Support Vector Machine with those features. The extracted features are related to the seam carving process. In addition, Local Binary Patterns transform is applied to the images before feature extraction to reveal the local artifacts caused by seam carving. Besides, the effect of seam carving in sub parts of the images is investigated. For this purpose, the images are divided into several stripes and each and every stripe is evaluated in terms of seam features. This evaluation has been improved the detection accuracies. Seam carving based resizing has been detected up to 99,9% in 30%scaled images by the developed method. The detection performance has been improved 20% on the average when compared with other methods in the literature. The detection performance is improved 26% in low scaling ratios like 3% and 6% which are harder to detect.