Date of Award

Spring 2017

Document Type


Degree Name

Doctor of Philosophy in Electrical Engineering - (Ph.D.)


Electrical and Computer Engineering

First Advisor

Yun Q. Shi

Second Advisor

Frank Y. Shih

Third Advisor

John D. Carpinelli

Fourth Advisor

Edwin Hou

Fifth Advisor

Xuan Liu


Digital images have been playing an important role in our daily life for the last several decades. Naturally, image editing technologies have been tremendously developed due to the increasing demands. As a result, digital images can be easily manipulated on a personal computer or even a cellphone for many purposes nowadays, so that the authenticity of digital images becomes an important issue. In this dissertation research, four machine learning based forensic methods are presented to detect one of the popular image editing techniques, called ‘seam carving’.

To reveal seam carving applied to uncompressed images from the perspective of energy distribution change, an energy based statistical model is proposed as the first work in this dissertation. Features measured global energy of images, remaining optimal seams, and noise level are extracted from four local derivative pattern (LDP) domains instead of from the original pixel domain to heighten the energy change caused by seam carving. A support vector machine (SVM) based classifier is employed to determine whether an image has been seam carved or not. In the second work, an advanced feature model is presented for seam carving detection by investigating the statistical variation among neighboring pixels. Comprised with three types of statistical features, i.e., LDP features, Markov features, and SPAM features, the powerful feature model significantly improved the state-of-the-art accuracy in detecting low carving rate seam carving. After the feature selection by utilizing SVM based recursive feature elimination (SVM-RFE), with a small amount of features selected from the proposed model the overall performance is further improved. Combining above mentioned two works, a hybrid feature model is then proposed as the third work to further boost the accuracy in detecting seam carving at low carving rate. The proposed model consists of two sets of features, which capture energy change and neighboring relationship variation respectively, achieves remarkable performance on revealing seam carving, especially low carving rate seam carving, in digital images. Besides these three hand crafted feature models, a deep convolutional neural network is designed for seam carving detection. It is the first work that successfully utilizes deep learning technology to solve this forensic problem. The experimental works demonstrate their much more improved performance in the cases where the amount of seam carving is not serious.

Although these four pieces of work move the seam carving detection ahead substantially, future research works with more advanced statistical model or deep neural network along this line are expected.