Araştırma Çıktıları | WoS | Scopus | TR-Dizin | PubMed
Permanent URI for this communityhttps://hdl.handle.net/20.500.14719/1741
Browse
6 results
Search Results
Publication Metadata only Deep Covariance Feature and CNN-based End-to-End Masked Face Recognition(IEEE, 2021) Junayed, Masum Shah; Sadeghzadeh, Arezoo; Islam, Md Baharul; Struc, V; Ivanovska, M; Bahcesehir UniversityWith the emergence of the global epidemic of COVID-19, face recognition systems have achieved much attention as contactless identity verification methods. However, covering a considerable part of the face by the mask poses severe challenges for conventional face recognition systems. This paper proposes an automated Masked Face Recognition (MFR) system based on the combination of a mask occlusion discarding technique and a deep-learning model. Initially, a pre-processing step is carried out in which the images pass three filters. Then, a Convolutional Neural Network (CNN) model is proposed to extract the features from unoccluded regions of the faces (i.e., eyes and forehead). These feature maps are employed to obtain covariance-based features. Two extra layers, i.e., Bitmap and Eigenvalue, are designed to reduce the dimension and concatenate these covariance feature matrices. The deep covariance features are quantized to codebooks combined based on Bag-of-Features (BoF) paradigm. Finally, a global histogram is created based on these codebooks and utilized for training an SVM classifier. The proposed method is trained and evaluated on Real-World-Masked-Face-Recognition-Dataset (RMFRD) and Simulated-Masked-Face-Recognition-Dataset (SMFRD) achieves an accuracy of 95.07% and 92.32%, respectively, showing its competitive performance compared to the state-of-the-art. Experimental results prove that our system has high robustness against noisy data and illumination variations.Publication Metadata only Denoising Diffusion Probabilistic Model for Retinal Image Generation and Segmentation(IEEE, 2023) Alimanov, Alnur; Islam, Md Baharul; Bahcesehir UniversityExperts use retinal images and vessel trees to detect and diagnose various eye, blood circulation, and brain-related diseases. However, manual segmentation of retinal images is a time-consuming process that requires high expertise and is difficult due to privacy issues. Many methods have been proposed to segment images, but the need for large retinal image datasets limits the performance of these methods. Several methods synthesize deep learning models based on Generative Adversarial Networks (GAN) to generate limited sample varieties. This paper proposes a novel Denoising Diffusion Probabilistic Model (DDPM) that outperformed GANs in image synthesis. We developed a Retinal Trees (ReTree) dataset consisting of retinal images, corresponding vessel trees, and a segmentation network based on DDPM trained with images from the ReTree dataset. In the first stage, we develop a two-stage DDPM that generates vessel trees from random numbers belonging to a standard normal distribution. Later, the model is guided to generate fundus images from given vessel trees and random distribution. The proposed dataset has been evaluated quantitatively and qualitatively. Quantitative evaluation metrics include Frechet Inception Distance (FID) score, Jaccard similarity coefficient, Cohen's kappa, Matthew's Correlation Coefficient (MCC), precision, recall, F1-score, and accuracy. We trained the vessel segmentation model with synthetic data to validate our dataset's efficiency and tested it on authentic data. Our developed dataset and source code is available at https://github.com/AAleka/retree.Publication Metadata only DeepPyNet: A Deep Feature Pyramid Network for Optical Flow Estimation(IEEE, 2021) Jeny, Afsana Ahsan; Islam, Md Baharul; Aydin, Tarkan; Cree, MJ; Bahcesehir UniversityRecent advances in optical flow prediction have been made possible by using feature pyramids and iterative refining. Though downsampling in feature pyramids may cause foreground items to merge with the background, the iterative processing could be incorrect in optical flow experiments. Particularly the outcomes of the movement of narrow and tiny objects can be more invisible in the flow scene. We introduce a novel method called DeepPyNet for optical flow estimation that includes feature extractor, multi-channel cost volume, and flow decoder. In this method, we propose a deep recurrent feature pyramid-based network for the end-to-end optical flow estimation. The feature extraction from each pixel of the feature map keeps essential information without modifying the feature receptive field. Then, a multi-scale 4 Dc orrelation volume is built from the visual similarity of each pair of pixels. Finally, we utilize the multi-scale correlation volumes to continuously update the flow field through an iterative recurrent method. Experimental results demonstrate that DeepPyNet significantly eliminates flow errors and provides state-of-the-art performance in various datasets. Moreover, DeepPyNet is less complex and uses only 6.1M parameters 81% and 35% smaller than the popular FlowNet and PWC-Net+, respectively.Publication Metadata only Towards Stereoscopic Video Deblurring Using Deep Convolutional Networks(SPRINGER INTERNATIONAL PUBLISHING AG, 2021) Imani, Hassan; Islam, Md Baharul; Bebis, G; Athitsos, V; Yan, T; Lau, M; Li, F; Shi, C; Yuan, X; Mousas, C; Bruder, G; Bahcesehir UniversityThese days stereoscopic cameras are commonly used in daily life, such as the new smartphones and emerging technologies. The quality of the stereo video can be affected by various factors (e.g., blur artifact due to camera/object motion). For solving this issue, several methods are proposed for monocular deblurring, and there are some limited proposed works for stereo content deblurring. This paper presents a novel stereoscopic video deblurring model considering the consecutive left and right video frames. To compensate for the motion in stereoscopic video, we feed consecutive frames from the previous and next frames to the 3D CNN networks, which can help for further deblurring. Also, our proposed model uses the stereoscopic other view information to help for deblurring. Specifically, to deblur the stereo frames, our model takes the left and right stereoscopic frames and some neighboring left and right frames as the inputs. Then, after compensation for the transformation between consecutive frames, a 3D Convolutional Neural Network (CNN) is applied to the left and right batches of frames to extract their features. This model consists of the modified 3D U-Net networks. To aggregate the left and right features, the Parallax Attention Module (PAM) is modified to fuse the left and right features and create the output deblurred frames. The experimental results on the recently proposed Stereo Blur dataset show that the proposed method can effectively deblur the blurry stereoscopic videos.Publication Metadata only PoseTED: A Novel Regression-Based Technique for Recognizing Multiple Pose Instances(SPRINGER INTERNATIONAL PUBLISHING AG, 2021) Jeny, Afsana Ahsan; Junayed, Masum Shah; Islam, Md Baharul; Bebis, G; Athitsos, V; Yan, T; Lau, M; Li, F; Shi, C; Yuan, X; Mousas, C; Bruder, G; Bahcesehir UniversityPose estimation for multiple people can be viewed as a hierarchical set predicting challenge. Algorithms are needed to classify all persons according to their physical components appropriately. Pose estimation methods are divided into two categories: (1) heatmap-based, (2) regression-based. Heatmap-based techniques are susceptible to various heuristic designs and are not end-to-end trainable, while regression-based methods involve fewer intermediary non-differentiable stages. This paper presents a novel regression-based multi-instance human pose recognition network called PoseTED. It utilizes the well-known object detector YOLOv4 for person detection, and the spatial transformer network (STN) used as a cropping filter. After that, we used a CNN-based backbone that extracts deep features and positional encoding with an encoder-decoder transformer applied for keypoint detection, solving the heuristic design problem before regression-based techniques and increasing overall performance. A prediction-based feed-forward network (FFN) is used to predict several key locations' posture as a group and display the body components as an output. Two available public datasets are tested in this experiment. Experimental results are shown on the COCO andMPII datasets, with an average precision (AP) of 73.7% on the COCO val. dataset, 72.7% on the COCO test dev. dataset, and 89.7% on the MPII datasets, respectively. These results are comparable to the state-of-the-art methods.Publication Metadata only An Effective Multi-Camera Dataset and Hybrid Feature Matcher for Real-Time Video Stitching(IEEE, 2021) Hosen, Md Imran; Islam, Md Baharul; Sadeghzadeh, Arezoo; Cree, MJ; Bahcesehir UniversityMulti-camera video stitching combines several videos captured by different cameras into a single video for a wide Field-of-View (FOV). In this paper, a novel dataset is developed for video stitching which consists of 30 video sets captured by four static cameras in various environmental scenarios. Then, a new video stitching method is proposed based on a hybrid matcher for stitching four videos with over 200 degrees FOV. The keypoints and descriptors are obtained by the scale-invariant feature transform (SIFT) and Root-SIFT, respectively. Then, these keypoint descriptors are matched by applying a hybrid matcher, a combination of Brute Force (BF), and Fast Linear Approximated Nearest Neighbours (FLANN) matchers. After geometrical verification and eliminating outlier matching points, one-time homography is estimated based on Random Sample Consensus (RANSAC). The proposed method is implemented and evaluated in different indoor/outdoor video settings. Experimental results demonstrate the capability, high accuracy, and robustness of the proposed method.