Development of image processing method for forgery detection

Nowadays, with the development of computer technologies, digital images are manipulated without leaving a clear trace, and simply forgerd images are created thanks to image processing software. With the use of these manipulated forgered images by malicious people in the media, it is a matter of debate whether the images are forgered or original (authentic). For this reason, there is a great need to determine the region where forgered images are manipulated in important fields such as politics, law and forensic medicine. Many studies have been carried out and various algorithms have been developed to detect the manipulated regions of forgered images. Today, deep learning methods are widely preferred by researchers because traditional methods for detecting image forgery are insufficient to detect forgered images obtained with advanced image processing software. These methods achieve higher success thanks to their automatic learning by acquiring complex features of images and their ability to classify forgered and original images. Deep learning enables the development of innovative image processing methods, especially convolutional neural networks. Therefore, convolutional neural networks can capture patterns in images more effectively and are more sensitive in detecting forgered images. In this study, a method that uses the advantages of traditional methods such as discrete cosine and discrete wavelet transform by combining the advantages of deep learning techniques is proposed to detect the manipulated region in forgered images. In the proposed method, an architecture is designed in which discrete wavelet transform and discrete cosine transforms are used in parallel with convolutional neural networks. In order to compare the success of the method, three different methods were applied, using only discrete cosine and convolutional neural networks, using only discrete wavelet transform and convolutional neural networks, and using only convolutional neural networks without using both transformations. A total of four different methods were tested on two different data sets and compared in terms of success metrics such as Accuracy, Precision, Recall, Dice Coefficient and F1 score. The results obtained from the benchmark clearly demonstrated the effectiveness and high classification accuracy of the proposed method.