Image splicing detection scheme using adaptive threshold mean ternary pattern descriptor

Abstract

The rapid growth of image editing applications has an impact on image forgery cases. Image forgery is a big challenge in authentic image identification. Images can be readily altered using post-processing effects, such as blurring shallow depth, JPEG compression, homogenous regions, and noise to forge the image. Besides, the process can be applied in the spliced image to produce a composite image. Thus, there is a need to develop a scheme of image forgery detection for image splicing. In this research, suitable features of the descriptors for the detection of spliced forgery are defined. These features will reduce the impact of blurring shallow depth, homogenous area, and noise attacks to improve the accuracy. Therefore, a technique to detect forgery at the image level of the image splicing was designed and developed. At this level, the technique involves four important steps. Firstly, convert colour image to three colour channels followed by partition of image into overlapping block and each block is partitioned into non-overlapping cells. Next, Adaptive Thresholding Mean Ternary Pattern Descriptor (ATMTP) is applied on each cell to produce six ATMTP codes and finally, the tested image is classified. In the next part of the scheme, detected forgery object in the spliced image involves five major steps. Initially, similarity among every neighbouring district is computed and the two most comparable areas are assembled together to the point that the entire picture turns into a single area. Secondly, merge similar regions according to specific state, which satisfies the condition of fewer than four pixels between similar regions that lead to obtaining the desired regions to represent objects that exist in the spliced image. Thirdly, select random blocks from the edge of the binary image based on the binary mask. Fourthly, for each block, the Gabor Filter feature is extracted to assess the edges extracted of the segmented image. Finally, the Support Vector Machine (SVM) is used to classify the images. Evaluation of the scheme was experimented using three sets of standard datasets, namely, the Institute of Automation, Chinese Academy of Sciences (CASIA) version TIDE 1.0 and 2.0, and Columbia University. The results showed that, the ATMTP achieved higher accuracy of 98.95%, 99.03% and 99.17% respectively for each set of datasets. Therefore, the findings of this research has proven the significant contribution of the scheme in improving image forgery detection. It is recommended that the scheme be further improved in the future by considering geometrical perspective.