Behavioural Study of Constrained Convolutional Neural Network on Image Splicing Classification of Various Datasets

Abstract

The Internet nowadays is exploded with digital images. However, it also has more forged images floating around and threatening the digital asset trustworthy, especially spliced images. Although there are many computational ways of detecting image splicing, the image splicing technique is advancing over time, driving feature specific algorithms to obsolete. Thus, more researchers nowadays are focusing on data-driven feature extraction. Convolutional Neural Network (CNN) is a Deep Learning algorithm well known for object detection, but it is ill-suited for manipulation feature extraction. In this research, a constrained convolution algorithm is to be injected into a simple CNN to study its performance in image splicing detection in a wide range of datasets. A strategic parameter tweaking can fit a wide range of spliced image datasets, but it is always biased to its train dataset. A cross-database splicing classification shows that it is ill-suited for generalizing a wide range of dataset evaluation.