Introduction

Digital photomontage refers to a paste-up produced by sticking together images regions using digital tools such as Photoshop. It is a popular technique used in visual content creation as well as malicious faking of images [Mitchell'94]. Several recent notorious cases in the news reporting involve the publication of faked images that are created by photomontage, e.g., an image showing the joint presence of Jane Fonda and John Kerry, and an image showing the prisoner abuse by British soldiers. To deter such malicious efforts, we need a general solution that enables any party receiving a digital image to verify the content authenticity without any overhead of watermark insertion or signature generation at the source. Such a technical is called blind, passive photomontage detection. Specifically, we are concerned with the blind detection of image splicing, a simple cut-and-paste of image regions without any post-processing, which lies in the core of the photomontage operation.

Recently, several research groups have started the investigation of this problem and proposed the use of image signal statistics and machine learning methods [Farid'99, Farid'03, NgChang'04a, NgChang'04b, NgChang'04c]. Such efforts have shown promising results and research directions. In order for researchers to compare and evaluate the pros and cons of different approaches, an open benchmark data set is needed.

Our objective is to compile a data set open to the research community so that new discovery and development of technologies can be expedited. The current data set is collected with sample diversity in mind. It has 933 authentic and 912 spliced image blocks of size 128 x 128 pixels. The image blocks are extracted from images in CalPhotos image set [CalPhotos'00]. The data set can be greatly improved in many ways, and should be considered as a preliminary effort addressing the increasingly important topic of benchmarking.

Design Criteria

We emphasize the following points while creating the data set.

• Content diversity: The data set contains 1845 image blocks (128 x 128 pixels) of diverse content extracted from the images of CalPhotos site as well as a small set of 10 images captured by ourselves.

• Balanced distribution: The numbers of the authentic and spliced images are approximately the same.

• Realistic operation: We simulate the process of creating spliced images with two types of operations, crop-and-paste along object boundaries vs. crop-and-paste of horizontal (or vertical) strips. Image objects and strips can be from the same image or two separate source images. Objects spliced together can be the same or different types, smooth or textured.

• Localized detection: We decompose the authentic as well as spliced images into separate local blocks of a fixed size (128 pixels x 128 pixels). The block is kept at a reasonable size to ensure that sufficiently accurate statistical features can be estimated using the empirical data of each block.

Copyright

The copyrights of the original images from the CalPhotos site are owned by the providers of the images. Information about the usage rights and other copyright issues can be found at http://elib.cs.berkeley.edu/photos/use.html

We thank the Berkeley Digital Library group for their generous support in making the original image set available for internal research. We are currently trying to seek the permissions from the prospective owners of the images for us to release the data set as a research benchmark. Status of such permissions and information about download procedures will be updated on the following site.

http://www.ee.columbia.edu/dvmm/downloads/AuthSplicedDataSet/AuthSplicedDataSet.htm

Structure of Data Set

The data set consists of 1845 image blocks of size 128x128 pixels.

There are two main categories of data set:

• (Au) = Authentic category: 933 image blocks

• (Sp) = Spliced category: 912 image blocks

The Authentic and Spliced categories are respectively subdivided in five subcategories:

• (T) = Image block with an entirely homogeneous textured region

• (S) = Image block with an entirely homogeneous smooth region

• (TS) = Image block with an object boundary between a textured region and a smooth region

• (TT) = Image block with an object boundary between two textured regions

• (SS) = Image block with an object boundary between two smooth regions

Then, the subcategories (TS), (TT) and (SS) is further subdivided into 3 sub-subcategories, according to the orientation of the object boundary:

• (V) = with vertical object boundary

• (H) = with horizontal object boundary

• (O) = other than (V) and (H)

Below are the some typical image blocks in each subcategory of the data set:

Figure 1 Typical image blocks in the data set

The following table shows that the number of image blocks in each subcategory:

Table 1: numbers of image blocks in different subcategories

Class Notation and File Structure

A sub-subcategory is named as a class. We denote a particular class with the following naming convention:

(main-categories)-(sub-categories)-(orientation)

(e.g.) A class under the authentic categories, with image blocks of having a vertical object boundary separating a textured and a smooth region is denoted as Au-TS-V class.

(e.g.) A class under the spliced categories, with image blocks of an entirely homogeneous textured region is denoted as Sp-T class.

Each class of image blocks is kept a directory with the same name as its class name within the zip file.

Image Sources

The original images used in this dataset consist of 312 images from CalPhotos collection and 10 images captured by us using a digital camera.

We considered images that come directly out of imaging devices as authentic and images resulted from splicing of image regions (of the same or different images) as spliced. More discussions and definitions of image authenticity can be found in [NgChang'04a].We made an assumption that images from the CalPhotos site are the original captured data from imaging devices (cameras or scanners) without editing or post-processing. However, this assumption has not been confirmed with the original image providers.

The following table lists the IDs of the images from the CalPhotos Images. CalPhotos images are originally organized in directory structure with the directory name being number code.

Table 2: IDs of images from the CalPhotos image collection.

To further diversify the image types, we use a digital camera to capture 10 additional images. Specifically, we mainly use this small set of images to populate more image blocks for Au-SS-V/O and Au-TS-H/V/O classes, as shown in Table 3 , where more detailed information about the 10 digital images are given Table 3 . Besides that, Figure 2 shows two examples of the digital images.

Table 3: Characteristics of the 10 digital images captured by using a digital camera

Data Set Production Procedure

In this section, we describe the procedures used in generating the authentic and spliced image blocks.

For the authentic category, the image blocks are cropped directly from the CalPhotos images and the images captured by the camera without any further manipulations.

For the spliced category, we first create composite images by object-based image splicing, i.e., cropping an object from an image and paste it to another image, without any post-processing such as edge smoothing and so on. Then, we cropped image blocks which contains part of the splicing boundary from the composite image as image blocks of the (Sp-TS-x), (Sp-TT-x) and (Sp-SS-x) classes, where x represents V, H or O. The above-mentioned cut and paste of image objects is performed using Adobe Photoshop 6.0 on a Windows XP. When cutting an object from an image, the outline of the object is traced using the Magnetic Lasso tool. Plain cut and paste commands are used for transforming the cut object to another image.

The image blocks of (Sp-T) and (Sp-S) classes (i.e., spliced image blocks of homogenous textured and smooth region) are created from the corresponding image blocks of (Au-T) and (Au-S) classes (i.e., authentic image blocks of homogenous textured and smooth region) respectively. Specifically, given an image block of (Au-T) or (Au-S) class, a corresponding image block for (Sp-T) or (Sp-S) class respectively is produced, by copying a vertical strip or a horizontal strip of 20 pixels wide (the decision on the strip orientation is random) from a randomly selected location, within the given image block of (Au-T) or (Au-S) class, and paste it to another randomly selected but different location within the same image (see Figure 3). The rationale for such way of producing images blocks of (Sp-T) and (Sp-S) classes from image blocks of (Au-T) and (Au-S) classes is for imitating the likely scenario of image forger patching up the void left behind after removing an object from an image. In this case, the void is likely to be covered up using small patches similar to the homogenous background. (Note that (Au-T) and (Au-S) classes consist of image blocks with homogenous texture or homogenous smooth region).

Figure 3
: The process for creating image blocks of (Sp-T) and (Sp-S) classes from the corresponding image blocks of (Au-T) and (Au-S) classes respectively: An example of creating the image block no. 10 of (Sp-T) from the image block no. 10 of (Au-T)

References

1. [NgChang'04c] Tian-Tsong Ng, Shih-Fu Chang. A Model for Image Splicing. In IEEE International Conference on Image Processing (ICIP), Singapore, October 2004.

2. [NgChang'04b] Tian-Tsong Ng, Shih-Fu Chang, Qibin Sun. Blind Detection of Photomontage Using Higher Order Statistics. In IEEE International Symposium on Circuits and Systems (ISCAS), Vancouver, Canada, May 2004.

3. [NgChang'04a] Tian-Tsong Ng, Shih-Fu Chang. Blind Detection of Digital Photomontage using Higher Order Statistics. ADVENT Technical Report #201-2004-1 Columbia University, June 2004.

4. [CalPhotos'00] CalPhotos, "A database of photos of plants, animals, habitats and other natural history subjects," http://elib.cs.berkeley.edu/photos/, 2000.

5. [Farid'99] H. Farid, "Detecting Digital Forgeries Using Bispectral Analysis," MIT AI Memo AIM-1657, MIT, 1999.

6. [Farid'03] H.. Farid, "A Picture Tells a Thousand Lies," New Scientist, vol. 179, pp. 38-41, 2003.

7. [Mitchell'94] W. J. Mitchell, "When Is Seeing Believing?," Scientific American, pp. 44-49, 1994.

Questions and Comments

Please forward any questions and comments

1. Tian-Tsong Ng

2. Shih-Fu Chang

Research Work which uses the Dataset

1. passive-blind image splicing detection project

Links

1. Detailed research description page for the Passive-blind Image Splicing Detection Project

2. Research summary page for other research projects in DVMM group

Acknowledgement

We sincerely appreciate Ginger Ogle of Berkeley Digital Library Project for the help in contacting the original photographers and transfering the image files.

For problems or questions regarding this web site contact The Web Master.
Last updated: June 12, 2003.