1 Introduction

The reliability of images is significant in numerous regions, such as scientific examination, criminal examination, observation frameworks, knowledge administrations, medicinal imaging and news coverage. Nowadays, in the computerized scenario, the information related to an image can be easily manipulated [29]. Digital image forensics is used to find out if an image has undergone any kind of forgery or not.

The digital investigation of an image provides the historic information of an image related to the camera or device properties used for shooting. Based on this information, one can find that the considered image is a doctored or not [29]. The image is compressed once during a shooting in the camera device, and if this image is further decompressed and resaved with different compression quality, the resultant image becomes double compressed. The first quantization matrix is always lost from the image metadata in the case of double compressed images. Therefore, estimation of the first quantization matrix from double compressed images is a challenging task for the researchers due to its importance in the digital investigation [20].

In an advanced examination, various acquisition devices and systems generally utilize the JPEG compression for images [7]. The issues faced during image examination can be divided into two broad categories [16], the first regarding the legitimacy of the visual document, and the other concerning the identification of the device which is used for the image acquisition.

Numerous methodologies are available to discover the JPEG images manipulations [30] and [28]. Different methodologies, as portrayed in [1] and [25], deal with the analysis of the statistical distribution of the DCT coefficient values. It is conceivable to figure out whether an image is doubly compressed or not by analyzing the related histogram [20] and [23]. The probability of each block of an image is calculated by a technique [10] which is being subjected to double compression.

To find out the difference in the images, a Markov random process is utilized in [4] by exploiting the SVM. The JPEG ghost technique [8] is used to detect the double compressed regions in an image and in [31] a periodic function detection strategy is used to identify the altered regions. A scheme is recommended in [12] to analyze the double compression using the same quantization matrix. Based on the estimated primary quantization matrix, a new technique is opined in [27] for the detection of double compressed JPEG images. Furthermore, a technique is depicted in [32] to detect the double JPEG compression by utilizing the same quantization matrix. In [33], a technique is proposed to detect the in-camera JPEG compression for double compressed images. Moreover, a robust technique is presented in [13] to detect the copy-move forgery in images under high JPEG compression artifacts. A novel technique for image analysis based on three visualization learning strategies is revealed in [34] to identify double JPEG compression clues. In the case of a duplicate operation, the image recompression by utilizing same quantization table is distinguished in [19], concluding that the grid of the pasted part has a significant probability that it is not aligned with the existing one. A new technique based on the compression artifacts presented in [3] focuses on discovering the traces caused by recompression.

The recreation of the first quantization matrix utilized by the acquisition device which is lost during double compression is important in all issues expressed previously. The model of the camera devices utilizing the same quantization tables is identified by estimating some part of the first quantization matrix. For the estimation of the first quantization matrix, in [20] the author exposes a few thoughts taking into account the conduct of normalized histograms.

The works in [35] and [2] additionally point out the image falsifications but do not provide a proper estimation of the first quantization matrix. In a splicing or a cloning operation, the DCT grid portion of an image pasted during duplication can or cannot be aligned to the original image which results in the occurrence of aligned and non-aligned traces. The first quantization matrix is essential among the parameters that are required to model the double compressed regions and accurately identify likelihood map [2]. A DCT coefficients histogram analysis is utilized in [24] to estimate the first quantization matrix from the double compressed JPEG image. In [9], a technique is proposed to estimate the first quantization matrix from double compressed JPEG images that cope with the case when the first quantization step is larger than the second. In the recent work [36], a novel distributed lossless coding technique is suggested for the hyperspectral images. Furthermore, a hyperspectral image compression and reconstruction technique based on the multi-dimensional or tensor data processing approach is presented in [37]. The hyperspectral images involve Distributive Source Coding (DSC) for the lossless compression. The artifacts left during the compression can be analyzed for the detection of forgery introduced in these images.

The work in this paper contributes towards rectifying the following problems:

  • All the existing techniques generally concentrate only on the detection of the double compressed region in an image. This provides a wide scope for the extraction of this double compressed part from the image.

  • Secondly, all the techniques in the related work are based on the estimation of first quantization matrix from the double compressed JPEG images. Therefore, estimation of first quantization matrix from partial double compressed JPEG images can increase the level of forensic investigation.

  • Also, most of the techniques ignore the effect of the error introduced during color conversions (YCbCr to RGB and vice versa), rounding and truncation of the values to eight-bit integers, etc. The first quantization matrix estimation accuracy can be improved by removing the effects of this error.

In this paper, a technique is proposed to estimate the first quantization matrix from partial double compressed JPEG images. In the first stage, a technique is proposed not only to detect but also to automatically isolate the doubly compressed region from the image. Therefore, this isolation of double compressed region solves the problem to estimate the first quantization matrix from partial double compressed JPEG images. The second stage analyzes the isolated doubly compressed region to estimate the first quantization matrix. In the latter stage, a filtering technique is proposed to optimize the performance of the algorithm by reducing errors. The proposed approach is solely dedicated to detect the region which is recompressed using a different quantization matrix.

The paper is organized as follows. In Section 2, the detailed background of JPEG compression along with its effects is discussed. It also includes the brief overview of the JPEG ghost detection technique. Section 3 covers the proposed scheme in detail. The experiment results are provided in Section 4 and conclusion is discussed in Section 5.

2 Background of JPEG compression

The first step in the JPEG compression is the Discrete Cosine Transform (DCT). The DCT of the image is performed by dividing the whole image into non-overlapping blocks of 8 × 8 blocks. The DCT is performed to segregate the high frequency components from the low frequency components of the image. Then a quantization is applied by using the 8 × 8 quantization matrix integer value for each DCT coefficient. An error is generated in this phase known as quantization error. This error is the key reason of loss of information in JPEG compressed images. After this, it is transformed into a data stream with the utilization of classic entropy coding [21]. The process of image compression is backtracked conversely while achieving the JPEG decompression as shown in Fig. 1. If the same scheme is employed on JPEG image, with a different quantization matrix, the result would yield a double compressed image.

Fig. 1
figure 1

Single JPEG compression and de-compression

Full size image

The quantization is followed by the rounding and truncating of real values to transform the range of integers into the range of [0; 255]. In this process, rounding values and truncation error are produced. In addition to these errors, another type of error occurs due to the conversion between RGB and YCbCr color spaces [21]. The error analysis can be performed on the 8-bit gray scale images. After double compression, the value of each double quantized coefficient c DQ can be modeled as [8]:

$$ {c}_{DQ}=\left[\left(\left[\frac{c}{q_1}\right]{q}_1+e\right)\frac{1}{q_2}\right] $$
(1)

where c denotes the single DCT coefficient value, q 1 and q 2 are the first and second quantization steps. The operator [.] indicates the rounding function, and e includes the error due to several operations like color space conversions, rounding, and truncation of the values to 8-bit integers, etc. To infer the value of q 1, a further compression is performed in a proper range with a novel quantization step q 3 and this error is computed by using an error function as follows [8]:

$$ {d}_e\left(c,{q}_1,{q}_2,{q}_3\right)=\left|\left[\left[\left[\frac{c}{q_1}\right]\frac{q_1}{q_2}\right]\frac{q_2}{q_3}\right]{q}_3-\left[\left[\frac{c}{q_1}\right]\frac{q_1}{q_2}\right]{q}_2\right| $$
(2)

If double compression is carried out on an image, only the last quantization steps are accessible but the first ones cannot be accessed because they are lost. The Fig. 2c indicates the DCT histogram of an image after double quantization. The primary quantization is carried out by taking a quantization step q 1 = 11 and then acquired values of DCT coefficients are de-quantized with same the quantization step. Conclusively, the values are again quantized using a quantization step q 2 = 7. Apparently, the distribution of the doubly quantized values contains periodic empty bins. This happens because during the second quantization, the coefficient values are re-distributed into more bins than in the first quantization [20].

Fig. 2
figure 2

(a) DCT histogram of an uncompressed image, (b) DCT histogram after the first compression with q 1 = 11 , (c) DCT histogram after second compression with q 2 = 7

Full size image

2.1 The JPEG ghost detection

The JPEG ghost detection technique [8, 38] has the capacity to localize the parts of an image which have gone through double compression. Consider a DCT coefficient c 1 which is quantized by an amount q 1. Then the resultant coefficient from first quantization is subsequently quantized second time by quantization step q 2 to result coefficient  c 2 . With the exception of q 2 = 1, the difference between c 1 and c 2 will be minimum when q 2 = q 1 and will increase as the difference between q 2 and q 1 increases. The JPEG ghost can be identified by considering each spatial frequency independently in each of the three luminance/color channels. However, various minima are possible when integer multiple quantization values are compared. The difference can be computed directly from the pixel values as follows:

$$ d\left(x,y,q\right)=\frac{1}{3}\sum_{i=1}^3{\left[f\left(x,y,i\right)-{f}_q\left(x,y,i\right)\right]}^2 $$
(3)

where f(x, y, i), i = 1, 2, 3, indicates each of three RGB color channels, and f q (.) is the result of compressing f(.) at quality q. There is some disparity in the difference images due to the underlying image content within and outside of the tampered region which could possibly be confound a forensic analysis as shown in Fig. 3.

Fig. 3
figure 3

Double compressed region detection through JPEG Ghost, along with the difference images corresponding to the different quality factor q

Full size image

3 Proposed scheme

The estimation of first quantization matrix from the partial double compressed JPEG images has significant importance in the forensic investigation. Therefore, a two-stage scheme is proposed in this paper for the estimation of first quantization matrix from the partial double compressed JPEG images. In the first stage, a technique is proposed to the automatic isolation of double compressed region from an image, and then this region is analyzed in the second stage to estimate the first quantization matrix. In the second stage, a filtering scheme is proposed to effectively reduce the effects of the error. The proposed scheme is for the scenarios when the image is recompressed with different quantization matrix than the first. The forgery introduced in an image through these partially double compressed JPEG images can also be analyzed.

In the proposed technique, the first stage as shown in Fig. 4 is the automatic isolation of double compressed region from an image by employing an enhanced JPEG Ghost detection technique. The conventional JPEG Ghost detection technique provides the gray scale difference image. The holes that occur in the gray scale difference image are then filled through morphology operation. Consequently, the adaptive median algorithm is applied after image complement which results in a binary image. The adaptive median algorithm classified the pixel values as noise by comparing each pixel value to its surrounding neighbored in the image. The pixel is considered as an impulse noise which is not structurally aligned with those pixels to which it is similar as well as which is different from a majority of its neighbors. These noise pixels are then replaced by the median pixel value of pixels in the neighborhood which have already passed the noise labeling test. The value of S max (maximum allowed size of the neighborhood) is adjusted according to the intensity value of double compressed part in difference image to bring it in the reasonable range of 21 ≤ S max ≤ 41. The given doctored JPEG image is then masked with the binary image which provides the masked image. The desired part is then cropped from the masked image. The whole first stage of the proposed scheme is depicted in Figs. 5 to 8.

Fig. 4
figure 4

Proposed scheme for the estimation of first quantization matrix from partial double compressed JPEG image

Full size image
Fig. 5
figure 5

(a) Partial double compressed image, Parrot (b) Difference image through JPEG ghost, (c) Image after morphology, adaptive filtering, and masking operation, (d) Isolated double compressed region of size 392 × 512 pixels

Full size image

The second stage analyzes the double compressed region to estimate the first quantization matrix as shown in Fig. 4. The resultant image from the first stage is properly cropped in the second stage and then the DCT coefficients of the image are extracted. The reason behind this cropping operation is that the extracted double compressed part would show some irregularities at the edges. These irregularities can be observed in Figs. 5d, 6d, 7d and 8d. These irregularities are present only at the edges and not at any other region. Thus, the error transferred to the second step can be minimized by proper cropping of the extracted double compressed part at the edges as shown in Fig. 9. The merits of proper cropping have already been discussed in [15, 20].

Fig. 6
figure 6

(a) Partial double compressed image, Roman (b) Difference image through JPEG ghost, (c) Image after morphology, adaptive filtering, and masking operation, (d) Isolated double compressed region of size 260 × 385 pixels

Full size image
Fig. 7
figure 7

(a) Partial double compressed image, Tower (b) Difference image through JPEG ghost, (c) Image after morphology, adaptive filtering, and masking operation, (d) Isolated double compressed region of size 82 × 205 pixels

Full size image
Fig. 8
figure 8

(a) Partial double compressed image, Wall (b) Difference image through JPEG ghost, (c) Image after morphology, adaptive filtering, and masking operation, (d) Isolated double compressed region of size 23 × 30 pixels

Full size image
Fig. 9
figure 9

(a) Resultant double compressed region from the first stage, Tower image (b) Image after proper cropping in the second stage

Full size image

Now the estimation of the first quantization matrix is carried out through the analysis of DCT coefficient histogram. The DCT histogram is filtered out through the proposed filtering scheme to reduce the effects of error e in (1). This filtering provides a set of filtered histograms. The rounding error e is modeled by approximating it as a Gaussian noise. This error proclaim itself by spreading peaks around the quantization step multiples. It will affect the second quantization step behavior along with the magnitude and statistics of the DCT coefficients. Due to this error two type of noises i.e. split and residual noise are encountered by the filtering strategy. The proposed histogram filtering scheme further reduces the effects of error e. By considering a set of first quantization steps in the range q 1i  ∈ {q 1min,  q 1min + 1, . .  . ., q 1max }, several filtering operations are then performed.

The properties of successive quantizations are exploited more efficiently by function (4) as compared to the error function (2). Therefore, by exploiting the q 1 localization property of (4), limited first quantization candidates (C s ) are selected. The function d out becomes nearly zero in the case q 3 = q 1i as shown in Fig. 10, when the filtering is done with the right first quantization step by using the formulation as follows:

$$ {d}_{out}\left(c,{q}_1,{q}_2,{q}_3\right)=\left|\left[\left[\left[\left[\frac{c}{q_1}\right]\frac{q_1}{q_2}\right]\frac{q_2}{q_3}\right]\frac{q_3}{q_2}\right]{q}_2-\left[\left[\frac{c}{q_1}\right]\frac{q_1}{q_2}\right]{q}_2\right| $$
(4)
Fig. 10
figure 10

Error function (4) for a AC coefficient with q 1 = 14, q 2 = 7 and q 3 ∈ {1, 2, . . ., 25}

Full size image

Each d out is evaluated for q 3 = q 1i by considering a single DCT frequency. If this value approaches to zero, it is included to the limited candidates set C s otherwise, discarded.

The double quantization process is simulated by considering the selected candidates. In this process, the first compression is performed by using selected candidates C s and the second compression with known quantization step and finally a histogram is selected that best exploits the original histogram. Therefore, the desired first quantization step is selected which corresponds to the final selected histogram from a pool of selected candidates C s .

3.1 Proposed DCT histogram filtering

Numerous methodologies [8, 9, 18, 23] generally do not consider the error e in (1) while employing the impacts of consecutive quantizations that take place after the de-quantizations. The performance of these methodologies reduces significantly when this source of error is neglected. But this simplification permits to handle the included mathematical equations easily. This error is introduced during the operations like color space conversions, rounding and truncation of the values. Based on the actual implementation, to examine each source of error separately is a challenging task. Therefore, the Gaussian distribution is used to model the overall effect of this error.

Several situations that can arise when an image is compressed more than once depend on the magnitude of the primary and secondary quantization steps. In the first scenario, the wrong bin DCT coefficient elements are propagated in the resultant histogram when a small perturbation occurs as shown in Fig. 11a. In the second scenario, the bin containing original information in the histogram can be equally split into two adjoining bins and out of these two bins, one is a wrong bin, as shown in Fig. 11b. This undesired situation arises when a primary quantization bin in position uq 1 falls exactly halfway between two back to back bins in position vq 2 and (v + 1)q 2 coming from the second quantization related as follows:

$$ u{q}_1=\frac{v{q}_2+\left(v+1\right){q}_2}{2},\kern1.5em u,v\ \epsilon\ {N}^{+} $$
(5)
Fig. 11
figure 11

(a) Represents the residual noise, (b) Represents split noise and (c) Represents the proposed split noise scenario

Full size image

Here a problem arises which has not been discussed previously, in which the bin from the second quantization (v + 1)q 2 becomes common to the two different cases of split noise, as shown in Fig. 11c. The bin in position uq 1 of the primary quantization is situated exactly halfway between two back to back bins in position vq 2 and (v + 1)q 2 coming from the second quantization. Similarly, the bin in position (u + 1)q 1 is positioned absolutely in the middle of two bins coming from the second quantization in position (v + 1)q 2 and (v + 2)q 2. Thus, to cope with this problem a filtering algorithm has been proposed. The proposed algorithm first finds out the wrong bins in the double quantized histogram with the help of (5) by considering the particular values of quantization steps q 1 and q 2. The identified wrong bins are then moved to the right locations according to the proposed algorithm as shown in Fig. 12b. On the other hand, the residual noise is removed by setting the proper threshold as shown in Fig. 12c.

Fig. 12
figure 12

(a) Double quantized DCT histogram, (b) DCT histogram after split noise removal through proposed filtering algorithm, (c) DCT histogram after residual noise removal

Full size image

4 Experiment results

The efficacy of the proposed approach is examined by conducting several tests on different datasets images. The JPEG encoding is used with standard JPEG quantization tables suggested by the Independent JPEG Group [14]. The database is generated by considering the Kodak lossless true color image (PhotoCD PCD0992) dataset [5] and UCID (v2) dataset [26] images. A set of 560 partial double compressed images is obtained from both the datasets considering the quality factors (QF1, QF2) in the range 50 to 100. The images in both the datasets are aligned. The supporting idea for the arrangement of images is to consider an uncompressed JPEG image and to double compress the regions of variable sizes in descending order. The outputs are reported with respect to the quality factors rather than the particular quantization steps. This step simplifies the analysis of the results, since a single quality factor parameter describes a quantization matrix with the 64 quantization steps that usually have different values corresponding to different frequencies as shown in Fig. 13 . Since the most of the information is carried by the lower frequency DCT coefficients, therefore the experimental analysis is based on the first 15 components. The proposed partial double compression detection technique does not provide satisfactory results for the images in which recompression is performed with a cropping attack and the same quantization matrix. Since, after the cropping attack, if the image is recompressed with same quantization matrix, will lead to the desynchronization of DCT blocks [11, 22]. The main problem is then to evaluate the desynchronization of DCT blocks introduced into the image. The proposed detection scheme is unable to detect this desynchronization. Thus, the case of recompression with same quantization matrix is not considered in this paper.

Fig. 13
figure 13

Quantization matrix for quality factor (QF) =50, according to the JPEG standard

Full size image

The performance analysis of the first stage of proposed technique has been done by considering the different detection parameters on two different datasets. The blocking artifacts introduced in double compressed regions are computed using different detection parameters as shown in Tables 1 and 2. It is clear from the Tables 1 and 2 that the values of blockiness signature measure (K F ) [6], gradient aware blockiness measure \( \left({B}_{gr}^{\lambda}\right) \) [39], and the calibrated feature (K L ) [17], vary according to the present artifacts and the size of the double compressed region. The discussed detection parameters measure smoothness of the image, where for smooth images their values approach to zero. Therefore, these parameters also confirm that the isolated region is actually double compressed. These blocking artifacts detection parameters measure the effectiveness of the first stage in automatic isolation of double compressed region. The Tables 1 and 2 also shows the average values of various detection parameters by considering a set of 560 images build from the Kodak lossless true color image (PhotoCD PCD0992 dataset) as well as the UCID (v2) dataset. Therefore, the average values of various detection parameters further confirm the efficacy of first stage.

Table 1 Performance analysis of the first stage of proposed scheme based on various detection parameters by considering the Kodak lossless true color image (PhotoCD PCD0992 dataset)
Full size table
Table 2 Performance analysis of the first stage of proposed scheme based on various detection parameters by considering the UCID (v2) dataset
Full size table

The percentage error of the first stage due to the edge irregularities of the isolated part increases with the reduction in the double compressed region size. The percentage error in isolation of double compressed region on the two different datasets i.e. Kodak lossless true color image (PhotoCD PCD0992) and UCID (v2) dataset is shown in Fig. 14. It is clear from the Fig. 14 that the error is larger in the case of UCID dataset due to their small image size which leads to the processing of fewer blocks as compared to the Kodak lossless true color image dataset.

Fig. 14
figure 14

Comparative analysis of percentage error (%) in the automatic isolation of the double compressed region through the first stage by considering Kodak lossless true color image (PhotoCD PCD0992) and UCID (v2) dataset

Full size image

The adaptive median algorithm is applied on the difference image generated from JPEG Ghost technique in which the boundary irregularities in the resultant binary image can be observed. The error due to the edge irregularities changes as the function of different JPEG compression factors. Tables 3 and 4 reports the average percentage of error in the first stage as a function of various JPEG compression factors on the Kodak lossless true color image (PhotoCD PCD0992 dataset) [5] and UCID (v2) dataset [26] respectively. It can be seen from the Tables 3 and 4 that the accuracy decreases in the case of UCID (v2) dataset as compared to the Kodak lossless true color image dataset due to their small image sizes.

Table 3 Average percentage of error in the first stage as a function of various JPEG compression factors on the Kodak lossless true color image (PhotoCD PCD0992 dataset)
Full size table
Table 4 Average percentage of error in the first stage as a function of various JPEG compression factors on the UCID (v2) dataset
Full size table

Tables 5 and 6 reports the average percentage error of the second stage by considering 560 partial double compressed images created from Kodak lossless true color image (PhotoCD PCD0992 dataset) as well as the UCID (v2) dataset. The percentage of error is calculated by estimating the q 1values at different quality factor values. Since the analysis is performed on the partial double compressed image datasets, therefore the isolated double compressed region is processed in the second stage to estimate the first quantization matrix whereas in the existing state-of-the-art techniques full double compressed images are used [2, 8, 9, 20, 24]. The DCT coefficients related to the low frequency in the DCT domain have small estimation error and do not essentially rely on the particular quality factor concerned with the first and second quantization. On the other hand, the estimation error for higher frequencies has significant correlation with the quality factors. The results obtained are usually better for higher QF1 and QF2 values as compare to the lower quantization. The comparative analysis of Tables 5 and 6 depicts that the percentage error in estimating the q 1 values at different quality factor corresponding to the first 15 DCT coefficients is more in case of UCID dataset as compare to the Kodak lossless true color image dataset due to the small size of images in UCID dataset.

Table 5 Average percentage of error in estimated q1 values at different quality factor corresponding to the first 15 DCT coefficients on the Kodak lossless true color image (PhotoCD PCD0992 dataset)
Full size table
Table 6 Average percentage of error in estimated q1 values at different quality factor corresponding to the first 15 DCT coefficients on the UCID (v2) dataset
Full size table

In order to study the effectiveness of the proposed approach, further analyzes have been led with respect to the specific DCT coefficient. The Tables 7 and 8 show the average percentage of error in the estimation of q 1 values that corresponds to the DCT coefficients on two different datasets. For the DCT coefficients corresponding to the high frequencies, the performance of the proposed scheme degrades but it provides high accuracy with an error less than 1.5% for the first 10 DCT coefficients as shown in Table 7.

Table 7 Average percentage of error in estimated q 1 values corresponding to the DCT coefficient in zig-zag order considering several state-of-the-art approaches on the Kodak lossless true color image (PhotoCD PCD0992 dataset)
Full size table
Table 8 Average percentage of error in estimated q 1 values corresponding to the DCT coefficient in zig-zag order considering several state-of-the-art approaches on the UCID (v2) dataset
Full size table

A comparative analysis of this second stage of proposed scheme is done with the various algorithms proposed in [2, 8, 9, 20]. The application of the second stage of the proposed scheme, when first 9 DCT coefficients are considered in a zig-zag manner, mixed results are obtained with few coefficients providing high error values. But, after the 9th DCT coefficient, the percentage error recorded is much low from the previously established techniques as shown in Table 7. If the Table 7 is closely examined, the percentage error can be conveniently justified. It can be inferred that the proposed scheme outperforms the method proposed by Bianchi [2] for all the considered 15 DCT coefficients. Whereas, the proposed scheme lag behind the schemes provided by Lukas [20] and Galvan [9] in only one DCT coefficient. Farid [8] surpass the proposed scheme for only two DCT coefficients.

The proposed filtering step considerably improves the performance of the second stage that copes with rounding and truncation error efficiently. The Tables 7 and 8 shows that the proposed scheme provides less percentage error as compared to the considered several state-of-the-art approaches in estimating the q 1 values corresponding to the DCT coefficient in zig-zag order on both datasets. It is clear from the Table 8 that the proposed scheme outperforms the existing techniques proposed by Lukas [20], Farid [8] and Bianchi [2] for all the considered 15 DCT coefficients. The technique provided by the Galvan [9] outperforms the proposed scheme for only two DCT coefficients.

The performances of the considered methods in the test on UCID (v2) dataset as shown in Table 8 are lower than on the Kodak lossless true color image dataset as shown in Table 7. The efficiency of considered schemes depends on the resolution of the image under analysis. The reliability of the analysis could be low as shown in Table 8, because of small size images in UCID (v2) dataset. Nevertheless, in today’s scenario working with small images is not so common.

5 Conclusions

The unethical use of digital images emerges the forensic investigation techniques to find out the authenticity of the digital images. In this paper, a technique is proposed to detect and analyze the partial doubly compressed JPEG image which leads to the estimation of the first quantization matrix. The proposed technique is based on the concept that to estimate the first quantization matrix from the partial double compressed image, it is desired to detect and isolate the double compressed region efficiently. The experimental results show that the first stage of the proposed scheme has satisfactory performances with an average percentage accuracy of 95.45% even when the detected double compressed regions are of a small size. The proposed filtering strategy increases the accuracy to estimate the first quantization matrix in the second stage. For the first 10 DCT coefficients, an error less than 1.5% has been recorded. The experiment results depict that the performance of the proposed approach is better as compared to the considered state-of-the-art techniques for the partial double compressed images based on the two different datasets. The evaluation is based on the partial double compressed images in which the recompression is performed with different quantization matrix. The work can be extended to create a detection methodology for the partial double compressed images, where recompression is performed with same quantization matrix. The proposed scheme can also be generalized from the color images to the hyperspectral images.