Abstract

Content-aware image retargeting has been investigated since the last decade as a paradigm of image modification for proper display on the different screen sizes. Modifications, such as seam carving or seam insertion, have been introduced to achieve aforesaid image retargeting. The changes in an image are not easily recognizable by human eyes. Inspired by the Blocking Artifact Characteristics Matrix (BACM), a method to detect tampers caused by seam modification on JPEG retargeted images without knowledge of the original image is proposed in this paper. In a BACM block matrix, we found that the original JPEG image demonstrates a regular symmetrical data, whereas the symmetrical data in a block reconstructed by seam modification is destroyed. Twenty-two features are proposed to train the data by using a Support Vector Machine (SVM) classification method. The experimental results clearly demonstrate that the proposed method provides a very high recognition rate for those JPEG retargeted images.


Index Terms—Image forensics, JPEG analysis, Seam-carving detection, Steganalysis features, Tamper detection

Authors
Kanoksak Wattanachote, Timothy K. Shih, Senior Member, IEEE, Wen-Lung Chang and Hon-Hang Chang
e-mail: kanoksak.wattanachote@gmail.com, timothykshih@gmail.com
National Central University,
No. 300, Jhongda Road, Jhongli City, Taoyuan County 32001, Taiwan (R.O.C.)
Source Code: Download all
Image Dataset

Seam Carving and Insertion of UCID Image at QF100:


Seam Carving and Insertion for UCUS Image at QF100:


To download our dataset, please contact to e-mail: kanoksak.wattanachote@gmail.com, timothykshih@gmail.com
1. Images Database: Download all
1) UCID Database (Zipped)
 2) UCID Database (1338 images)
 3) UCUS Database (Zipped)
 4) UCUS Database (1009 images)
Experimental Results

     This research introduces a seam tamper detection method for the JPEG retargeted images. The workflow of system is shown in Fig. 4. The starting images are prepared in accordance with Sarkar et al.’s approach. The starting images are prepared as concluded below.
      1. Compress the images in JPEG at QF75.
      2. Decompress the images before seam modifications.
     3. Retarget the images by using seam modifications for each different tampering rates and classify into different dataset, namely 1%, 2%, 5%, 10%, 20%, 30%, 50%, and mixed set.
     4. Compress the retargeted images to be JPEG at QF100 before passing through the detection process as shown in Fig. 4. The accuracy results are subsequently observed from the last step in Fig. 4, and recorded to demonstrate as in Tables.


      The results in Table II and III demonstrate that the accuracy for testing by low tampering rate models such 1% or 2% are higher than that test by high tampering rate models such 50%. The tamper detection results for both seam carving and seam insertion obtained using UCID images are not different and similar to that obtained using UCUS images as demonstrated in perspective views in Fig. 11 and Fig. 12.


     

     

     The highest average accuracy for seam carving detection was found in the experiment with QF100, whereas the lowest was found in the experiment with QF50. Besides, the highest average accuracy for seam insertion detection was found in the experiment with QF10, whereas the lowest average accuracy was found in the experiment with QF75.

  
  

    Table XII (a) shows Sarkar et al.’s results and (b) demonstrates Wei et al.’s results. Table XII (c)-(e) demonstrates the results derived from our proposed method obtained using our two databases. The results show that the accuracy values by our proposed method are higher than that derived from Sarkar et al.’s method. In average, the accuracy by our proposed method is also higher than the accuracy by Wei et al.’s method. The experiments in Table XII (c)-(e) aimed to validate the dependence of the detection accuracy and the image databases. The results show that the detection accuracy and the image databases are independent since the results in Table XII (e) is similar to the results in Table XII (c) and (d). The detection accuracy by our proposed method in average is around 98-99%, higher than that obtained by Sarkar et al.’s and Wei et al.’s methods.

      The experimental results in Table XIII show that the average accuracy of seam insertion detection obtained by our proposed method is also higher than that obtained by Sarkar et al.’s method. That means the symmetric phenomenon in blocking effects is evident for the compressed image. Seam modifications both carving and insertion can destroy that symmetric phenomenon.

      We have also experimented with uncompressed images obtained using the images from UCID database. The UCID images are retargeted and subsequently passed through the detection processes as described in Fig. 4, without compression process. The experimental results in Table XIV show that test by mixed model can lead the detection accuracy in average around 67-69%, both in seam reduction and seam insertion detection. This means the symmetric phenomenon in blocking effects is not evident for uncompressed image.

      Fig. 21 demonstrates the average accuracy of cross validation obtained using UCID and UCUS databases, from Table II to Table XI. The cross validation sets are conducted in 8 different tampering rates consist of 1%, 2%, 5%, 10%, 20%, 30%, 50%, and mixed rates; and 5 different tampering rates consist of 10%, 20%, 30%, 50%, and mixed rates. The two charts in Fig. 21 demonstrate that the average accuracy of those two cross validation sets are not different. The highest average accuracy for seam carving detection was found in the experiment with QF100, whereas the lowest was found in the experiment with QF50. Besides, the highest average accuracy for seam insertion detection was found in the experiment with QF10, whereas the lowest average accuracy was found in the experiment with QF75.
      We conducted an experiment for tamper images at unknown QF, by combining the compressed images at different QF for tampering in each dataset. The tamper images at unknown QF are equally obtained using UCID database. The results are shown in Table XV with the accuracy charts in perspective view demonstrated in Fig. 22 (a) and (b). The detection of tamper images at unknown QF can be implicitly explained as an average accuracy of the detection obtained from the results in experiment at QF10, QF20, QF50, QF75, and QF100, as shown in Fig. 24 (a) and (b). The highest average accuracy for seam carving detection was found in QF100 test, whereas the lowest was found in QF50 test. Besides, the highest average accuracy for seam insertion detection was found in QF10 test, whereas the lowest average accuracy was found in QF75 test.


Experimental Results of QFs Cross Validation







     For the detection by cross QF with similar factors (e.g., QF10 and QF20), in small seam removal rate (1% - 10%), it’s possible to use cross QF detection, such as using QF20 for testing QF10. Whereas, the cross detection in seam insertion is possible to use QF10, QF20, and QF50 for testing QF10 and QF20 in every tamper rate, since the detection accuracy results of QF10 and QF20 tested by QF10, QF20, and QF50 are significantly higher than that testing by other QF. Obviously, the 50% tamper rate in seam insertion datasets, cross detections by different QF can be effectively implemented. The reason is, by changing 50% of seams in an image always make big different between the original and the tampered image, which can also be recognized by SVM classification method.
    Additionally, the average of accuracy results in Table XV in our revised paper is quite similar to the average of accuracy derived from Table 1-7 (in response the comments).

Experimental Results of Seam Tampered Detection in Removing Object Purpose

     We cite the recommended paper in Section V. A. accordingly, as reference [31] in our revised paper. The proposed seam-carving approach in [31] is to remove a small object from the image by preserving the image size, whereas the purpose of seam-carving in our experiment is to resize the image (as demonstrated in Fig. 24, in our revised paper). To preserve the image size after tampering in [31], both seam removal and seam insertion are implemented. We conducted an experiment for tamper detection based on the new reference. We downloaded the seam-carving tool (“Seam Carving GUI” - SeamCarvingGui-Mac-1.11.tar.gz) as recommended in [31] from http://code.google.com/p/seam-carving-gui/. We used this tool to remove the object or unwanted region from our sample JPEG images for testing. A sample result is demonstrated in Figure 2 (above). We use 10 untouched images and 10 tampered images for each dataset. The target regions for our three experiments are marked by red color as demonstrated in the middle columns. The tamper detection is done by the SVM classification method as used in our study. The results are reported below (for two datsets and three samples/dataset).


     The detection is conducted by the SVM classification method using the 22-features data for train and test, as demonstrated below.

         Phoenix:windows Phoenix$ svm-train BACM10.txt
         *.*
         optimization finished, #iter = 21
         nu = 0.886715
         obj = -11.782504, rho = 0.071702
         nSV = 20, nBSV = 15
         Total nSV = 20
         Phoenix:windows Phoenix$ svm-predict BACM10.txt BACM10.txt.model testBACM10.txt;
         Accuracy = 100% (20/20) (classification)

     For the SMALL (small region removal) dataset, the tamper detection accuracy is 100%.


     The detection is conducted by the SVM classification method using the 22-features data for train and test, as demonstrated below.

         Phoenix:windows Phoenix$ svm-train BACM10-3.txt
         *
         optimization finished, #iter = 18
         nu = 0.958450
         obj = -13.354675, rho = 0.074884
         nSV = 20, nBSV = 16
         Total nSV = 20
         Phoenix:windows Phoenix$ svm-predict BACM10-3.txt BACM10-3.txt.model testBACM10-3.txt;
         Accuracy = 95% (19/20) (classification)

  For the tamper detection accuracy of the SMALLEST (SMALL object is larger than SMALLEST object) dataset, the tamper detection accuracy is 95%. The SVM for this dataset could not totally classify the difference of the untouched images and seam tamper images, since only a very small number of seams were modified.

  We compare the target region removal size in our experiments with the removed object size obtained using [31]. The target region size for removal in our SMALL dataset is close to the removed object size in [31]. The accuracy for tamper detection for this kind of seam-carving images by using our proposed method is 100%, higher than that using the method proposed in [31]. The target region size for this kind of tamper JPEG image is an important variable for seam tamper detection accuracy. Our seam tamper detection method for JPEG image performs higher than the method proposed in [31].

Experimental Results of Forward and Backward Energy (Source codes)

     We conduct small experiment to find the relationship of seam tamper with different energy functions used for seam carving. We use the UCID dataset consists of 500 sample images for this experiment. The results show that no significantly difference of seam tamper detection accuracy for those two approaches. Logically, since the symmetry phenomena in BACM will always be destroyed with the same seam tamper rate. Our proposed method will always sensitive with the BACM block.

       Copyright 2015 - MINE Lab, National Central University