The proposed system is put into place to fix the issues with the current systems. The image that would be used to communicate between users was taken as an input at first. The following processes, which are illustrated in the following figure.1, are used to process that image with the aid of the root sifting and feature point matching algorithms.
A) Feature Extraction
Reducing the contrast threshold and resizing the input image are two straightforward but effective ways to extract a significant number of critical points from the first step of feature extraction, even from smooth and minute regions. A threshold is used for the rejection of unstable extrema having low contrast values. The challenge of creating a sufficient number of key points cannot be solved by decreasing the contrast threshold alone. When copy-move forgery is carried out on tiny areas. The supplied image is resized by a factor of two as a complementary technique. The proposed system generates key points based on SIFT and ROOFSIFT. The SIFT technique is separated into four phases:
-
Identifying Candidate key points.
-
Refining of key points based on contrast and edge thresholds.
-
Each key point’s dominant orientation assignment
-
Generation of feature descriptor
At the start of phase 1, candidate Key points are selected on various scales. Consecutive Gaussian-blurred photos are formed from an input image I by repeatedly convolving it using Gaussian filters at various sizes. Then, local extrema inside a 3 x 3 x 3 Difference of Gaussians (DoG) zone are selected as the prospective critical spots. All potential crucial locations are further adjusted in phase 2 using a contrast and edge threshold. The SIFT algorithm's removal of unstable extrema depends on this strategy. Phase 3 involves assigning a dominant orientation to each important point that survives in order to guarantee rotation invariance.
Aggregating gradient orientation information, a histogram for orientation is produced from points in a local window centered on the SIFT key point. In step 4, a 128-dimensional descriptor is produced by coding the surrounding information in a restricted area using the SIFT key point. Simply lowering the contrast threshold won't solve the issue of getting enough key points when copy-move forgery is used on tiny areas. Before computing SIFT key points, enlarge the input picture by a factor of S. Then comparing histograms, the Hellinger kernel or the chi-squared distance generally outperforms the Euclidean distance. Most of the SIFT articles describe comparing descriptors using the Euclidean distance, SIFT is still a histogram itself. As a result, rather than employing a separate measure to compare SIFT descriptors, the 128-dim descriptor produced by SIFT can be adjusted. RootSIFT, proposed by Arandjelovic et al[4], is a simple algebraic modification to the SIFT descriptor that allows SIFT descriptors to be "compared" using a Hellinger kernel. The following is a basic approach for extending SIFT to RootSIFT:
Feature Extraction Algorithm :
Input : Image / Frame
Output resized and feature extracted image
Step 1: initialize the SIFT feature extractor and then Compute SIFT descriptors.
Step 2: if there are no key points or descriptors, return an empty tuple.
Step 3: Apply the Hellinger kernel by first L1-normalizing and taking the square root.
Step 4: Then the vectors are L2-normalized and return a tuple of the key points and descriptors.
B) Matching of Feature Points
The matching of feature points in the copy-move forgery detection scenario searches the picture for comparable local areas. Scale clustering is a matching technique that calculates the distance between each key point in a cluster to identify the keypoint k's matched correspondence (C1, C2, or C3). The more crucial sites there were, the greater the computing overhead would be. An effective matching approach is even more crucial in this design since the feature extraction process generates a high number of key points. To hasten the matching procedure, a method known as overlapping grey level clustering is applied. The suggested method can be used to carry out the matching procedure much more effectively without having to eliminate initial valid matches. The key points are divided into clusters that overlap. The nearby clusters, in particular, will share a portion of the key points. Initially, based on the RGB values [0, 1, ….., 255] is evenly split into M intersecting sub-levels, each having a length of c1 and an intersecting length of c2. It's important to note that none of the matched pairs in P are in any particular order. Because the matching procedure ignores the location information, it is unable to provide a constant matching direction.
C) Localization of Forged Region
Localization identifies duplicated spots in sparse fields in the forgery detection scenario. When it comes to locating the fraudulent portions, keypoint-based picture copy-move forgery detection[29] approaches have two problems. 2) Because there is typically no matching order in matched pairs, the forged points and their original counterparts are not broken up during the matching process, even when many clones are made, the homography is typically distinct and the percentage of duplicated areas is unknown. The framework of the suggested forgery localization approach is made up of four parts.
Step 1) First, isolated matched pairings are eliminated or each matched pair is removed to lower the false alarm rate. (m, m’ ) Є P, where P is defined as the set containing all matched pairs, let Nm and Nm’ be the numbers of matching points with location distances to m and m’ lesser than some threshold T.Here matched pairs that satisfying Eq. (1) is discarded.
After this step, M is a new set consisting of the matched survived pairs.
Step 2)
For homographic estimation, a subset of related pairs from two nearby local areas will be used to estimate an affine matrix. After selecting a matched pair at random, all matched key points around the selected matched pair are recorded. All the related pairs that are near to the given matched pairs are combined to form a new set. It's worth noting that all of the related pairs in the set have the same matching order. Because all of the related pairings in the collection are made up of two contiguous local areas, it's safe to conclude that they all follow the same homography. The homography between correspondence of the matched pairs in the set is estimated using the RANSAC algorithm [16].
Step 3)
Validation of homography and selection of inliers based on prevailing orientation. To discard inaccurate estimation from RANSAC, by accentuating the prevailing orientation associated with each keypoint, a homography validation, and inlier selection strategy has been developed. The offset of the dominating orientations for each inlier must be consistent with the predicted affine homography. Correctness of the estimated homography is validated by a function f defined as
For an adequately computed homography and correctly matched pair, the function f is near zero. To avoid the occurrence of inliers that are incorrectly accepted by chance, we accept the homography if and only if the new set matches over 90% of the time.
Step 4)
Localization of forgery using the scale and RGB information. There are two phases to this approach i.e i) construction of suspicious regions and (ii) refinement of suspicious regions. A local suspected region is created in the first step based on the scale information of each inlier. In the second phase, the suspicious regions are refined by exploiting the color information. This localization algorithm operates iteratively and will be repeated for N iterations. After finishing all the iterations, the detected forgery regions can be generated.