Browsing by Author "Tora, Hakan"
Now showing 1 - 7 of 7
Results Per Page
Sort Options
Item A GENERALIZTION OF ARNOLD'S CAT MAP AND FRACTION BASED EMBEDDING IN IMAGE STEGANOGRAPHY(2022-02-15) Buker, Mohamed; Tora, Hakan; Gökçay, ErhanThe rapid development of data communication, and the increased amount of information that are communicated via networks, make it very important to find new ways to protect exchanged information. Encryption is one of the most widely used methods nowadays in this area. Steganography is a recent field of research in which the communicated information is being invisible to anyone rather than being only encrypted. The idea behind steganography is to hide the existence of information itself. As long as a third party knew there were information, whether encrypted or not encrypted, the information will be at risk. In this thesis, we present a steganographic model with two levels of security. First, the secret image is scrambled using our Generalized Arnold Cat Map (ACM). Then, the scrambled image is embedded into another image using our Fraction Based Embedding Technique (FBE) in the transform domain using both Discrete Wavelet Transform (DWT) and Lifted Wavelet Transform (LWT). The efficiency of our model was tested on benchmark color images. Peak Signal-to-Noise Ratio (PSNR), Mean Square Error (MSE), Structural Similarity (SSIM) and correlation values are calculated. Results show that our Generalized ACM is more robust compared to standard and modified versions of ACM. At the same time, results of our new FBE technique performs better than those of other techniques regarding to PSNR and MSE values.Article A Novel Data Encryption Method Using an Interlaced Chaotic Transform(Expert Systems with Applications, 2024-03-01) Gökçay, Erhan; Tora, HakanWe present a novel data encryption approach that utilizes a cascaded chaotic map application. The chaotic map used in both permutation and diffusion is Arnold’s Cat Map (ACM), where the transformation is periodic and the encrypted data can be recovered. The original format of ACM is a two-dimensional mapping, and therefore it is suitable to randomize the pixel locations in an image. Since the values of pixels stay intact during the transformation, the process cannot encrypt an image, and known-text attacks can be used to get back the transformation matrix. The proposed approach uses ACM to shuffle the positions and values of two-dimensional data in an interlaced and nested process. This combination extends the period of the transformation, which is significantly longer than the period of the initial transformation. Furthermore, the nested process's possible combinations vastly expand the key space. At the same time, the interlaced pixel and value transformation makes the encryption highly resistant to any known-text attacks. The encrypted data passes all random-data tests proposed by the National Institute of Standards and Technology. Any type of data, including ASCII text, can be encrypted so long as it can be rearranged into a two-dimensional format.Item FACIAL EXPRESSION IDENTIFICATION USING TEXTURE AND SHAPE BASED FEATURES(2016-12-17) Gül, Nuray; Tora, HakanRecently, facial expression recognition (FER) systems have a significant role to play in the human-computer interaction (HCI) applications. In many existing systems, either the features of the whole face or the combination of the features extracted from some regions of face are used while defining an emotion. This study suggests using just one appropriate region for every single expression identification to demonstrate what is the effect of these regions on the feelings separately. In the proposed design, it’s aimed to identify Surprised and Happy emotions by using shape features of mouth region on the other hand the texture features of the eye region is used for Fear, Anger and Disgust emotions. Therefore, Fourier Descriptors (FD) and Local Binary Patterns (LBP) are extracted as feature vectors and these features are classified by using neural networks (NN). The system was trained on the Extended Cohn-Kanade Dataset (CK+) and achieved accuracy rate is almost 88.9% for the overall system.Item GASTRIC CANCER CLASSIFICATION USING DEEP CONVOLUTIONAL NEURAL NETWORK(2022-01-20) Jebur, Saif Salam; Tora, HakanIn this thesis, several pre-trained CNN and our CNN structure presented to automatic detection of early gastric cancer in endoscopic images. In the first stage, the transfer learning using two types normal and cancer of image datasets, the pre-trained networks executed for gastric cancer detection using MATLAB 2018. Then, the obtained results compared with each other and discussed in detail form. In the second stage, new structure proposed by using CNN. The proposed structure consists from 8 layers with SoftMax classifier. The extracted high-level features by convolutional layers classified by SoftMax in last layer. The proposed network presented 99.88% which is high result when compared with numerous performed pre-trained networks. Furthermore, the proposed network presented remarkable execution time when compared with several transfer learning techniques.Item IMPACT ASSESSMENT OF SEA LEVEL RISING (SLR) FOR KARASU COASTAL AREA USING GEOGRAPHIC INFORMATION SYSTEM (GIS) MODELING(2022-02-25) Eliawa, Ali; Tora, Hakan; Numanoğlu Genç, AslıSea Level Rise (SLR) due to global warming is becoming more imperative issue to coastal zones. In this thesis, an overall analysis has been made to assess the vulnerability of coastal area in Karasu region in Turkey. Based on predictions of sea level rise scenarios of 1m, 2m, and 3m, inundation levels were visualized using Digital Elevation Model (DEM). The eight-side rule algorithm was applied through Geographic Information System (GIS) by using a high resolution DEM data that are generated by utilizing eleven 1:5000 scale topographic maps issued by the National Land Survey of Turkey. The outcomes of GIS-based inundation maps indicate that 1.43 % of the total land area or 0.79 km2 , 6.16% of the total area or 3.4 km2 or, and 30.08 % of the total land area or 16.6 km2 are flooded for 1m, 2m, and 3m SLR scenarios respectively. Risk maps have shown that water bodies and beach areas have a higher risk for 1 m. scenario and for the 3 m. scenario urban areas, water bodies and the beach areas have higher risk. From the combined hazard map with vulnerability data, it is seen that estuarine areas at the west and east of Karasu region have a medium vulnerability while the inland behind the coastal zone has a low vulnerability. These results can provide a primary assessment data for Karasu region to the decision makers to enhance policies and planning of land use.Item NEURAL NETWORK BASED FEATURE EXTRACTION FOR HANDWRITTEN DIGIT RECOGNITION(2017-01-07) Günler Pirim, Mine Altınay; Tora, Hakan; Öztoprak, KasımIn this dissertation, it is proposed that hidden layer output weights of semi-trained neural network to be used as feature vectors. In pattern recognition neural network is a training algorithm which provides classification. In this thesis in addition to this fact, it has been shown that semi-trained neural network can be used as a tool to extract hidden layer output vectors that are used as features of the image. The system is mainly composed of three steps: preprocessor, feature extractor, and classifier. Only the classifier layer differs for each experiment, the other two layers are used as default for all experiments. Support vector machine, neural network, and Euclidean distance classifiers are utilized. The experiments were conducted on MNIST and USPS benchmark datasets to evaluate the performance of the proposed approach.Item PALM PRINT IDENTIFICATION(2022-02-24) Jebriel, Belal Ali Mesbah; Tora, HakanThis thesis explores the appropriateness of identifying palm prints through a standard database and a classifier. This study uses two sets of databases, CASIA and IIT, which contain left hand and right hand images. The features of the local binary pattern (LBP) and histogram of oriented gradients (HOG) are extracted from the images by MATLAB. Training and testing sets are created from these features. A multilayer neural network and support vector machines (SVM) with two separate kernels, linear and quadratic, are trained and tested on the selected databases. The chosen features are empirically compared with one another. Better results have been accomplished in HOG for both classifiers. In addition, the performance of the classifiers are evaluated. It has been observed that the neural network achieves better results than SVM for LBP features of both datasets. On the other hand, for HOG features, they do not display many advantages over one another.