Performance Appraise of Machine Learning Classifiers in Image Splicing Detection using Thepade’s Sorted Block Truncation Coding

Abstract

Image Splicing is known as a conventional type of digital image manipulation. It is one such type of tampering; also called as image composition. A spliced (or composite) image is usually created by copying and pasting portions of the image onto the same or another image. Spliced image detection mainly deals with finding similarity present in an image and establishing a relationship between authentic image parts and pasted portions of the image. With the increasing popularity and usage of easily available image editing technologies, even for people with minimal expertise it has become much easier to edit image data. Hence splicing is becoming sophisticated day by day making it difficult to detect with naked eyes. Due to the advent of social media and other platforms these spliced images can be circulated in faster ways among users of those platforms and hence it becomes necessary to come up with methods of spliced image detection. This paper proposes use of Thepade’s Sorted BTC, various Machine Learning classifiers for splicing detection. Here TSTBTC-Nary is explored with values of n as 2, 4, 6,…16,18 attempted on some machine learning classifiers (BayesNet, NaiveBayes, Logistic, Simple Logistic, SVM, JRip, PART, J48, LMT) for various performance metrics. After validation on 3 benchmark datasets CASIA V1, Columbia and Columbia-Uncompressed, LMT classifier performs better closely followed by Simple Logistic and J48. Better image splicing capabilities are observed with TSTBTC 16-ary closely followed by 18-ary and 14-ary.