Scopus İndeksli Yayınlar Koleksiyonu / Scopus Indexed Publications Collection
Permanent URI for this collectionhttps://hdl.handle.net/11147/7148
Browse
6 results
Search Results
Now showing 1 - 6 of 6
Conference Object Citation - WoS: 3Citation - Scopus: 7A Taxonomic Survey of Model Extraction Attacks(IEEE, 2023) Genç, Didem; Özuysal, Mustafa; Tomur, EmrahA model extraction attack aims to clone a machine learning target model deployed in the cloud solely by querying the target in a black-box manner. Once a clone is obtained it is possible to launch further attacks with the aid of the local model. In this survey, we analyze existing approaches and present a taxonomic overview of this field based on several important aspects that affect attack efficiency and performance. We present both early works and recently explored directions. We conclude with an analysis of future directions based on recent developments in machine learning methodology.Article Citation - WoS: 2Citation - Scopus: 2Improving Outdoor Plane Estimation Without Manual Supervision(Springer, 2022) Uzyıldırım, Furkan Eren; Özuysal, MustafaRecently, great progress has been made in the automatic detection and segmentation of planar regions from monocular images of indoor scenes. This has been achieved thanks to the development of convolutional neural network architectures for the task and the availability of large amounts of training data usually obtained with the help of active depth sensors. Unfortunately, it is much harder to obtain large image sets outdoors partly due to limited range of active sensors. Therefore, there is a need to develop techniques that transfer features learned from the indoor dataset to segmentation of outdoor images. We propose such an approach that does not require manual annotations on the outdoor datasets. Instead, we exploit a network trained on indoor images and an automatically reconstructed point cloud to estimate the training ground truth on the outdoor images in an energy minimization framework. We show that the resulting ground truth estimate is good enough to improve the network weights. Moreover, the process can be repeated multiple times to further improve plane detection and segmentation accuracy on monocular images of outdoor scenes.Article Citation - WoS: 55Citation - Scopus: 56Evaluation of an Artificial Intelligence System for Diagnosing Scaphoid Fracture on Direct Radiography(Springer Verlag, 2020) Özkaya, Emre; Topal, Fatih Esad; Bulut, Tuğrul; Gürsoy, Merve; Özuysal, Mustafa; Karakaya, ZeynepPurpose The aim of this study is to determine the diagnostic performance of artificial intelligence with the use of convolutional neural networks (CNN) for detecting scaphoid fractures on anteroposterior wrist radiographs. The performance of the deep learning algorithm was also compared with that of the emergency department (ED) physician and two orthopaedic specialists (less experienced and experienced in the hand surgery). Methods A total 390 patients with AP wrist radiographs were included in the study. The presence/absence of the fracture on radiographs was confirmed via CT. The diagnostic performance of the CNN, ED physician and two orthopaedic specialists (less experienced and experienced) as measured by AUC, sensitivity, specificity, F-Score and Youden index, to detect scaphoid fractures was evaluated and compared between the groups. Results The CNN had 76% sensitivity and 92% specificity, 0.840 AUC, 0.680 Youden index and 0.826Fscore values in identifying scaphoid fractures. The experienced orthopaedic specialist had the best diagnostic performance according to AUC. While CNN's performance was similar to a less experienced orthopaedic specialist, it was better than the ED physician. Conclusion The deep learning algorithm has the potential to be used for diagnosing scaphoid fractures on radiographs. Artificial intelligence can be useful for scaphoid fracture diagnosis particularly in the absence of an experienced orthopedist or hand surgeon.Conference Object Citation - WoS: 1Citation - Scopus: 2Scene text localization using keypoints(Institute of Electrical and Electronics Engineers Inc., 2015) Erdoğmuş, Nesli; Özuysal, MustafaScene text localization and recognition (also known as text localization and recognition in real-world images, nature scene OCR or text-in-the-wild problem) is an open problem, attracting increasing interest from researchers. In this paper, we address the localization issue and leave the recognition part out of its scope. For the purpose of scene text localization, Scale-Invariant Feature Transform (SIFT) keypoints are extracted from the images and classified as text and non-text. Subsequently, the text keypoints are utilized to compute the bounding boxes around text regions. The proposed technique is tested on the database of ICDAR 2013 Robust Reading Competition-Challenge 2 and the experimental results are reported in detail. Although the idea introduced here is still at its infancy, it is observed to achieve remarkable results and due to the fact that there is a large room for improvement, it is found to be promising.Conference Object A Detailed Analysis of Mser and Fast Repeatibility(Institute of Electrical and Electronics Engineers Inc., 2015) Uzyıldırım, Furkan Eren; Köksal, Ali; Özuysal, MustafaThis paper investigates the relationship between the MSER and FAST repeatability and changes in various camera parameters. By employing a realistic view synthesis methodology, it is possible to observe the effect of small parameter changes on the repeatability. Furthermore, for the analysis of MSER repeatability, a convex hull approach is proposed instead of fitting ellipses to the MSER region. This yields a better approximation to the MSER region without significantly increasing computation time.Conference Object Deep Convolutional Neural Networks for Viability Analysis Directly From Cell Holograms Captured Using Lensless Holographic Microscopy(The Chemical and Biological Microsystems Society (CBMS), 2019) Delikoyun, Kerem; Çine, Ersin; Anıl İnevi, Müge; Özçivici, Engin; Özuysal, Mustafa; Tekin, Hüseyin CumhurCell viability analysis is one of the most widely used protocols in the fields of biomedical sciences. Traditional methods are prone to human error and require high-cost and bulky instrumentations. Lensless digital inline holographic microscopy (LDIHM) offers low-cost and high resolution imaging. However, recorded holograms should be digitally reconstructed to obtain real images, which requires intense computational work. We introduce a deep transfer learning-based cell viability classification method that directly processes the hologram without reconstruction. This new model is only trained once and viability of each cell can be predicted from its hologram. © 2019 CBMS-0001.