In deep learning, small input shifts or translations can cause dramatic changes in the output. This is because of the fact that commonly used down-sampling techniques such as max pooling, strided convolution, and average pooling ignore the sampling theorem. We have demonstrated that the cycle-spinning (CS) signal processing technique can be used before down-sampling in deep learning to increase accuracy without introducing any extra learnable parameters. The proposed method can be applied to different algorithms such as GAN, classification, and object detection.

Date: 29.11.2022 / 13:00 Place: METU Informatics Institute

Developing a tool having the capability of predicting future AAA growth rate is important in terms of surgical planning and patient management. In this study, 106 CT scan images from 25 Korean AAA patients were retrospectively obtained. 21 geometrical measurements derived from scans, their growth rates, and their pairwise correlations were analyzed, and the predictability of the growth for high-risk aneurysms was attempted to enhance. Furthermore, the prediction model was built specifically on patient characteristics using the various geometrical measurements enhanced the prediction capability of a measurement at any time-point, along with an evaluation of the associated uncertainty.

Date: 02.03.2020 / 12:40 Place: Conference Hall-1

In this thesis, we offer a context-aware security model extending the role-based access control model in order to prevent relay attacks in NFC enabled mobile devices with both theoretical and practical approach. Within this study, we identify possible vulnerabilities and requirements then design the model. Parallel to conceptual design, we also developed a complete test-bed to deploy the model on it. Finally, we verified the model theoretically and practically.

Date: 27.01.2020 / 16:00 Place: Conference Hall-1

Vector Dynamic Time Warping (VDTW), a novel multi-year classification approach based on warping of angular distances between phenological vectors was developed. The results prove that the proposed VDTW method is robust to temporal and spectral variations compensating for different farming practices, climate and atmospheric effects, and measurement errors between years. We also describe a method for determining the most discriminative time window that allows high classification accuracies with limited data. We carried out tests of our approach with Landsat 8 time-series imagery from years 2013 to 2015 for classification of corn and cotton in the Harran Plain, and corn, cotton, and soybean in the Bismil Plain of Southeastern Turkey. In addition, VDTW was tested with corn and soybean in Kansas, the US for 2017 and 2018 with the Harmonized Landsat Sentinel data.

Date: 15.01.2020 / 14:00 Place: Conference Hall-1

Face-to-face conversation implies that interaction should be characterized as an inherently multimodal phenomenon involving both verbal and nonverbal signals. Gaze is a nonverbal cue that plays a key role in achieving social goals during the course of conversation. The purpose of this study is twofold: (i)to examine gaze behavior (i.e., aversion and gaze on face) and relations between gaze and speech in face to face interaction, (ii)to construct computational models to predict gaze behavior using high-level speech features.

Date: 13.01.2020 / 10:30 Place: Conference Hall-1

In this study we present a brain model for decoding mental states that are captured using functional brain imaging (fMRI). We postulate that, the human brain processes information coming from the senses using specialized brain regions and the brain combines the activity of the specialized regions to come up with a coherent mental state. We model the postulated pattern of information processing in the brain as follows: First, we propose to capture the activity of specialized brain regions using homogenous voxel (volumetric pixel) groups: Supervoxels. Second, we combine the activity of supervoxels to decode the overall mental state using classifier ensembles: Brain Region Ensembles (BRE).

We test our model in three distinct fMRI datasets, where our model performs better, in terms of accuracy of classification of mental states, than the widely used brain decoding methods that rely on voxel selection. Also, we present how supervoxels can be used for the localization of the brain regions that are effective in discriminating the mental states under consideration regarding fMRI experiments.

Date: 07.10.2019 14:30 Place: Conference Hall 1

Medical ultrasound provides various diagnostic advantages. If the image quality is improved, it will be beneficial for clinical usage.  In this thesis, the aim is to improve image quality and reduce speckle in B-mode ultrasound images. Both single and multi-frame, in-plane, freehand, 2D scan data was used for this purpose. Non-rigid registration, Bayesian restoration and super-resolution methods, along with a detailed study on statistical modelling of the speckle was employed. The results were compared to widely accepted image filtering methods. In terms of objective evaluation metrics and in image quality assesment on radiology experts, the proposed  methods performed better.

Date: 24.07.2019 Place: Conference Hall 1