Digital Image Forgery Identification Using Motion Blur Variations as Clue

Transcription

1 Digital Image Forgery Identification Using Motion Blur Variations as Clue P. M. Birajdar*, N. G. Dharashive** Abstract: Fake images have become common in society today. In all forms of media one can easily find forged images used to sensationalize news, spread political misinformation etc. Claims of image tampering are common while its truthfulness argument. As the trustworthiness of images suffers, it is necessary to devise techniques in order to verify their authenticity. Authentication can be determined using active or passive detection techniques.a variety oftechniques have been proposed for splicing detection. One of the novel techniques is developed for the passive detection of image forgeries utilizing motion blur as cue in order to detect spliced object. Block wise Motion blur estimation within image assists to identify probable splicing. Keywords: Image forgery, Blur, Motion blur estimation, Splicing detection. Author Details : P. M. Birajdar Department Computer Science and Engineering, M.S. Bidve Engineering College, Latur, Maharashtra, India ppornima09@gmail.com, Contact Number: INTRODUCTION A picture may worth a thousand words but, alongside, it may have scores of interpretations. Images and videos can be altered with a variety of common editing tools. [1] Some of the manipulated images have even received awards for their so called originality. Owing to such sophisticated digital image/video editing software tools, the establishment of the authenticity of an image has become a challenging task, encompassing a variety of issues due to this; there is a huge question mark over the use of multimedia data as evidence in the courts of law. Digital image forensics is a field that analyses images of a particular scenario N. G. Dharashive Department Computer Science and Engineering, M.S. Bidve Engineering College, Latur, Maharashtra, India Contact No ID-ngdharashive@gmail.com to establish reliability and authenticity. [1] It is fast becoming as a popular field because of its potential applications in many domains, such as intelligence, sports, legal services, news reporting, medical imaging and insurance claim investigations. For example, the imagein figure 1taken from [2] shows image forgery; on July 9, 2008 Iran announced it had successfully test-fired missiles with a range of 1,200 miles. An image of the test-firing is shown in figure1(a), showing the launch of four missiles at an undisclosed location in the Iranian desert, was made available by Sepah News, the media arm of Iran s Revolutionary Guard (IRG). The image was used by many media outlets. However, a photo of the test-firing

2 showing only three missiles launching as is figure1(b), emerged the same day. It first appeared on the Iranian news website Jamejam. Closer examination revealed that the first photo had been doctored. One of the missiles (second from the right) had been digitally created by cutting-and-pasting together elements of the other missiles. This was apparently done by the IRG in an attempt to conceal the failure of one of the missiles to launch. (a) (b) Fig.1. Image forgery(a) Doctored Image (b) Original Image Recently researchers are giving significant attention to the digital image forgery identification. At least two trends account for this: the first accepting digital image as official document has become a common practice, and the second the availability of low cost technology in which the image could be easily manipulated. Copy and move forgery, image splicing, image retouching are categories of image tampering techniques. Techniques from each of these categories can be implemented with two approaches: Active and Passive approach. 1. Active Approach: in this prior information is crucial and which is concerned with data hiding technique such as digital watermark, digital signature. Digital watermarking is used for image authentication. 2. Passive Approach: In contrast, passive approaches for image forensic operate in the lack of watermark or signature. No visual clues and prior information that indicate tampering. Such forgery detection is challenging problem. To address the possible image forgery, many tools are available.[3] Set of image forensic tools are categorized in to techniques based on pixel, format, Camera Response Function (CRF), physics, geometric. Pixel based technique detects statistical anomalies introduced at pixel level, format based techniques considers the factors like lossy compression scheme, JPEG quantization, Double JPEG, JPEG blocking, camera-based techniques that exploit artifacts introduced by the camera lens, sensor etc, physically based techniques that explicitly model and detect anomalies in the threedimensional interaction between physical objects, light, and geometricbased techniques that make measurements of objects in the world and their positions relative to the camera

3 In this paper we focus on digital image forgery identification using variation in motion blur as clue. The organization of this paper is as follows. Section 2.1 presents an overview of the cause of motion blur in an image, section 2.2 gives background information about the blur estimation process, and our proposed method of forgery identification is in Section 3. Experimental results are addressed in Section RELATED WORK 2.1. Overview of Motion Blur Blur is opacity in an image, so it can be used as cue to detect doctored image. Figure 2 shows that the turbulent medium like atmosphere, camera shake or fast moving objects introduces blur while taking an image. An image can have different types of blurs such as defocus blur, motion Blur, out-of-focus blur. Camera shake, high shutter speed or fast moving objects during image capturing causes blur with motion in an image. Primarily considering parameters of motion blur, artificial blur can be generated. Artificial Blur operation is used often to reduce the degree of discontinuity to hide splicing effects or to give real situation effect. [4]Artificial blurring changes the blur consistency pattern of original image and identifying blur inconsistencies in whole forged image can aid to detect image forgeries. Fig.2. Blur cause For example, the image in Figure 3 taken from[2], [4] this photograph was commonly spread via , purportedly having been obtained from a camera found in the debris of the World Trade Center buildings after the attacks of September 11, The forth coming air plane in the background appears to imply that this image was captured few seconds before the impact. Though, this photograph is proved to beforged. There are many cues within this image that help decide that it is a doctored image. A prior information may be employed to prove that this image is unauthentic. For example, geographical knowledge or information about the type of airplane involved in the attacks can be used to dismiss the claim that it s the original one. Even in the absence of such knowledge, as the camera is focused on the person, the airplane should have appeared blurred in the image due to its speed. The complete absence of motion blur of airplane in this image indicates a possible forgery [4]. Fig 3. Forged Tourist Guy image

4 2.2. Overview of motion blur estimation techniques An overview of image forgery detection technique is as shown in figure 4, blur of input image is estimated, and the variation in blur metric is used to segment the forged region. The blur estimation is possible with different techniques like Radon transformation, Hough transformation, Harr-wavelet edge based blur estimation, using cepstrum. To extract motion blur parameters, problem image is converted into frequency domain. Periodic patterns in frequency domain estimates the motion blur parameters. [5] The point spread function (P SF) has two important parameters, i.e. motion direction and motion length. The estimation of these parameters is very important for image forgery detection. Fig.4. Overview of Image Forgery Detection 3. PROPOSED IMAGE FORGERY IDENTIFICATION TECHNIQUE We propose a technique to identify image forgeries using motion blur estimates. Blur estimates are first computed for the given image. Every blur estimate contains the parameters; blur magnitude, and blur direction. Our technique segments the image based on the blur estimate, resulting the regions that have variation in blur. However the technique highlights such regions in both authentic and forged problem images, it is preferred to expose the possible forgery in inconsistent blurred regions, such as spliced objects with artificial blur close to original image blur because introducing an artificial blur is very common method used to decrease level of discontinuity or to hide splicing effects or to give real situation effects which is difficult to detect. The proposed forgery detection technique makes use of a Hough transform. [5] Hough transform is applied to find global patterns as lines, circles, and ellipses in an image in a parameter space. It is especially useful in line detection because lines can be easily detected as points in Hough transform space, based on the polar representation of line; ρ=x cosθ+ y sinθ (1) Where (x, y) represent Cartesian coordinates of a point on the line; θ is the angle between the x-axis and the perpendicular line to the given line going across the origin of the graph, ρ is the length of the perpendicular. Thus, a pair of coordinates ( ρ, θ) can describe the line. The blur direction is obtained by locating the maximum value of Hough transform of edge map of image

5 Y (θ,ρ) θ ρ π π Ө X Image Hough Domain Fig 5: Hough Transform Algorithm: Image forgery detection based on Motion Blur Discrimination Step 1: Input: Convert blurred RGB problem image to gray level image. Step 2: Convolution: Perform convolution over the image to remove boundary artifacts. Step 3:Edge Detection: Find the edge map of result of step 2. Step 4: Subdivision: Divide image into overlapping blocks. Step 5: Blur Estimation : Blockwise findblur parameters magnitude and direction using hough transformation, and the peak in Hough transform (the maximum value) which is perpendicular to the motion blur angle. Step 6: Interpolation: Using bicubic interpolation, blur estimates are up sampled to the size of image I. Step 7: Segmentation: Segment forged region of image if any, by using discrepancies in motion blur. A. Block-Level Analysis Given an image with artificially motion-blurred spliced region, H.Ji and C.Liu[6] presented a spectral analysis of image gradients, which leads to a better configuration for identifying the blurring kernel of more general motion types (uniform velocity motion, accelerated motion and vibration) and ρ hybrid Fourier-Radon transform to estimate the parameters of the blurring kernel with improved robustness to noise over available techniques. But it is not possible to extract multiple blur models over the whole image from its gradients, especially when the blurs are quite similar to each other. Hence, local level blur estimation is used for different blur models. The image I is divided into b overlapping blocks of size M X N and the motion blur estimate for each block is calculated. Motion blur can be estimated at a number of points, as opposed to just a single estimate for the entire image, giving improved resolution, is the major benefit of image subdivision. B. Blockwise Blur Estimation For the case of uniform motion blur, [4] the blurring process is modeled as the convolution of a sharp image with a blurring kernel: I(x,y)=(H*P)(x,y)+N(x,y) (2) wherei is blurred image, H is the original sharp image, P is the blurring kernel, and N is the noise present. x and y are the pixel coordinates. For a horizontal uniform velocity motion blur, the blurring kernel P u can be modeled as P u =1/L[1,1,.,1] 1XL, where L is the length of the kernel. Edge map of blurred image is detected using canny edge detector or sobel operator. Peak values of blockwise hough transformation gives blur direction as shown in figure 6. The value corresponding to the maximum value of the hough transform is the angle perpendicular to the blur direction. True blur direction is given by 90-θ

6 offered by bicubic interpolation over the other two is most clearly visible. (a) (b) Fig.6. (a) Spliced image (b) image showing blockwise blur direction with inconsistencies. C. Interpolation Bicubic interpolation is used to upsample the motion blur estimates to image size, which gives blur estimates at each pixel. The accuracy of the estimate depends on the amount of up sampling done. [4]Bicubic interpolation provides better results than nearest neighbor (which gives a blocky segmentation), and bilinear interpolation for which the segmentation has a few jagged edges that could be adequate for certain applications. Fig.7 shows an example of the segmentation outputs for various interpolations. The circles are examples of regions where the improvement Fig. 7. Segmentation outputs for various interpolation schemes applied to blurestimates [4]. (a) Nearest neighbor. (b) Bilinear. (c) Bicubic. D. Segmentation We then segment the image into two regions that show different motion blurs. This is done by adaptive thres holding as well as by performing morphological operations on blur direction. This method also provides an effectiveness metric which is used to discard images which show consistent directions and/or magnitudes in their motion blur estimates. The result of segmenting the magnitude and direction of the estimates provides us with an indication of regions with dissimilar motion blur. Fig.8. Segmentation: Spliced object in image of fig.6. (a)

7 (a) (b) Fig.9 (a) Doctored photo: India s Prime Minister Narendra Modi looking at flooded city of Chennai.(b) Blur Direction variations of forged image. (c) Segmented region (d)spliced region highlighted by red lines Figure 9 shows that, a doctored photo released by the some unidentified source has shown Prime Minister Narendra Modi surveying severe flood affected area of Chennai from a helicopter. [7] It is presented as Modi looking through the round window of a helicopter, through which a clear view of waterlogged buildings were visible. A similar photograph, but with a blurred view through the window, was posted on Modi's personal Twitter feed. Commentators on social media accused the PIB of digitally manipulating its version of the picture. The proposed technique based on discrepancies in motion blur direction in this paper has detected the spliced region in the forged image. 4. RESULTS (c) (d) Fig. 10 Images dataset Proposed system is tested on image dataset of 15 forged images, shown in figure10, where blur is introduced due to camera shake, and motion blur. We spliced different objects on the blurred backgrounds of the images using image Adobe

8 photoshop editor. The results of detection of spliced regions for some image are shown in figure11.in the case of both the ideal segmentation and the proposed segmentation based on blur direction, the boundary of the extracted region is used as the boundary resulted bythe segmentation for evaluation purpose. (a) (b) Fig. 11Identification of spliced region.(a) Forged image with segmentation output(b) Ideal Segmentation. block size for processing. However, the lower block size gives better segmentation than larger block size. Graphical representation of average cost for different block size and for different block overlap is as shown in figure 12 and figure 13 respectively. Block Size Segmentation cost 20 X X X X Table 1 : Segmentation cost for different block size. Block Overlap Segmentation cost Table 2: Average Segmentation cost for different block overlap. In order to evaluate the efficiency of our method we employ computational cost measure for blur estimation and segmentation of overlapping blocks. Table 1 shows the segmentation cost with respect to the block size and table 2 with respect to the block overlap. We calculate the motion blur estimates for block sizes of 20 20, 30 30, 40 40, and pixels with overlaps of 5, 10, 15, and 20 pixels. The minimum average cost is noted among that of 15 different images for each of 16 possible combinations of overlap and block size. It is observed that as block size increases, the cost of estimation is reduces, hence we prefer maximum Fig. 12 Average segmentation cost for different block size with same overlap

9 Fig. 13 Average segmentation cost for different block size with different overlap CONCLUSIONS In this paper we have presented an approach which uses dissimilarities in motion blur for image forgery detection. To accomplish this, primarily motion blur is estimated using different estimation metrics. Block level analysis of motion blur helps to get fine results. Considering the estimated blur as clue, forged region in image is segmented. An issue for forgery detection arises if original image carries multiple motion blurs. Large block size processing results in improper segmentation of forged region as compare to small block size processing. Limitations of proposed technique can be a challenge for researchers to explore new ideas and provide new solutions to the problem. REFERENCES : 1. T Qazi, K Hayat, I Khan, Survey on blind image forgery detection in IET Image Process., 2013, Vol. 7, Iss. 7, pp Museum of Hoaxes.[Online] Available : 3. HanyFarid, Image Forgery Detection in IEEE Signal Processing Magzine, March PravinKakar, SudhaNatrajan, Wee Ser, Exposing Digital Image Forgeries by detecting discrepancies in Motion Blur in IEEE Transactions On Multimedia,June 2011,Vol.13,No ShamikTiwari, V. P. Shukla, Ajay Kr. Singh, Certain Investigations on Motion Blur Detection and Estimation 6. H. Ji and C. Liu, Motion blur identification from image gradients, in Proc. IEEE Conf. Computer Vision and Pattern Recognition, 2008, pp [online]: flooded-criticismafter-photoshop-disaster-drags-modi-missionmuck?utm_medium=twitter&utm_source=twitterfeed