Sharif Digital Repository

This paper is focused on video event recognition based on frame level CNN descriptors. Using transfer learning, the image trained descriptors are applied to the video domain to make event recognition feasible in scenarios with limited computational resources. After fine-tuning of the existing Convolutional Neural Network (CNN) concept score extractors, pre-trained on ImageNet, the output descriptors of the different fully connected layers are employed as frame descriptors. The resulting descriptors are hierarchically post-processed and combined with novel and efficient pooling and normalization methods. As major contributions of this work to the video event recognition, we present a... 

We address blind extraction of spread spectrum data embedded in a digital image. The multicarrier iterative generalized least-squares (M-IGLS) data extraction algorithm is the basis of our study and a new algorithm based on M-IGLS is introduced in which both computational complexity and data extraction accuracy are enhanced. In the new method named multicarrier consecutive iterative generalized least-squares (M-CIGLS) algorithm, neither the cover image nor the spread spectrum carriers are required for the hidden data extraction. Additionally, knowledge of the number of carriers is no more needed. Simulation results show the ability of this method to extract data with a superior performance... 

A new method for exploiting the semantic hierarchical structure of visual concepts in video event recognition task is proposed in this paper. The visual concepts are detected using the readily available Convolutional Neural Network (CNN) structures which make the recognition system extremely efficient in cases with limited hardware resources. The employed CNNs assign scores to each of the predetermined visual concepts in each video frame and the resulting concept scores are fed to the proposed hierarchical post-processing scheme. Our post-processing module takes advantage of the semantic hierarchy of the concepts to enhance the recognition accuracy of event recognition. The hierarchical... 

This paper is focused on video event recognition based on frame level convolutional neural network (CNN) descriptors. Using transfer learning, the image trained descriptors are applied to the video domain to make event recognition feasible in scenarios with limited computational resources. After fine-tuning of the existing CNN concept score extractors, pretrained on ImageNet, the output descriptors of the different fully connected layers are employed as frame descriptors. The resulting descriptors are hierarchically postprocessed and combined with novel and efficient pooling and normalization methods. As major contributions of this paper to the video event recognition, we present a... 

This paper is focused on video event recognition based on frame level convolutional neural network (CNN) descriptors. Using transfer learning, the image trained descriptors are applied to the video domain to make event recognition feasible in scenarios with limited computational resources. After fine-tuning of the existing CNN concept score extractors, pretrained on ImageNet, the output descriptors of the different fully connected layers are employed as frame descriptors. The resulting descriptors are hierarchically postprocessed and combined with novel and efficient pooling and normalization methods. As major contributions of this paper to the video event recognition, we present a... 

Classification of video events based on frame-level descriptors is a common approach to video recognition. In the meanwhile, proper encoding of the frame-level descriptors is vital to the whole event classification procedure. While there are some pretty efficient video descriptor encoding methods, temporal ordering of the descriptors is often ignored in these encoding algorithms. In this paper, we show that by taking into account the temporal inter-frame dependencies and tracking the chronological order of video sub-events, accuracy of event recognition is further improved. First, the frame-level descriptors are extracted using convolutional neural networks (CNNs) pre-trained on ImageNet,... 

Diffusion weighted imaging (DWI) is a non-invasive method for investigating the brain white matter structure. It can be used to evaluate fiber bundles in the brain. However, clinical acquisitions are often low resolution. This paper proposes a method for improving the resolution using sparse representation. In this method a non-diffusion weighted image (bO) is utilized to learn the patches and then diffusion weighted images are reconstructed based on the trained dictionary. Our method is compared with bilinear, nearest neighbor and bicubic interpolation methods. The proposed method shows improvement in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity (SSIM)  

Diffusion tensor magnetic resonance imaging (DTMRI) is a non-invasive method for investigating the brain white matter structure. It can be used to evaluate fiber bundles in the brain but in the regions with crossing fibers, it fails. To resolve this problem, high angular resolution diffusion imaging (HARDI) with a large number of diffusion encoding directions is used and for reconstruction, the Q-ball method is applied. In this method, orientation distribution function (ODF) of fibers can be calculated. Mathematical models play a crucial role in the field of ODF. For instance, in registering Q-ball images for applications like group analysis or atlas construction, one needs to interpolate... 

Diffusion Tensor Imaging (DTI) is a common method for the investigation of brain white matter. In this method, it is assumed that diffusion of water molecules is Gaussian and so, it fails in fiber crossings where this assumption does not hold. High Angular Resolution Diffusion Imaging (HARDI) allows more accurate investigation of microstructures of the brain white matter; it can present fiber crossing in each voxel. HARDI contains complex orientation information of the fibers. Therefore, registration of these images is more complicated than the scalar images. In this paper, we propose a HARDI registration algorithm based on the feature vectors that are extracted from the Orientation... 

In this paper, we compare registration results obtained using different diffusion maps extracted from diffusion tensor imaging (DTI). Fractional Anisotropy (FA) and Ellipsoidal Area Ratio (EAR) are two diffusion maps (indices) that may be used for image registration. First, we use FA maps to find deformation matrix and register diffusion weighted images. Then, we use EAR maps and finally we use both of FA and EAR maps to register diffusion weighted images. The difference between FA values before deformation and after registration using the FA alone or EAR alone has a median of 0.57 and using both of them has a median of 0.29. Therefore, the results of registration using both of the FA and... 

Background: Diffusion weighted imaging (DWI) is a non-invasive method for investigating the brain white matter structure and can be used to evaluate fiber bundles. However, due to practical constraints, DWI data acquired in clinics are low resolution. New method: This paper proposes a method for interpolation of orientation distribution functions (ODFs). To this end, fuzzy clustering is applied to segment ODFs based on the principal diffusion directions (PDDs). Next, a cluster is modeled by a tensor so that an ODF is represented by a mixture of tensors. For interpolation, each tensor is rotated separately. Results: The method is applied on the synthetic and real DWI data of control and... 

Background and objective Registration is a critical step in group analysis of diffusion weighted images (DWI). Image registration is also necessary for construction of white matter atlases that can be used to identify white matter changes. A challenge in the registration of DWI is that the orientation of the fiber bundles should be considered in the process, making their registration more challenging than that of the scalar images. Most of the current registration methods use a model of diffusion profile, limiting the method to the used model. Methods We propose a model-independent method for DWI registration. The proposed method uses a multi-level free-form deformation (FFD), a sparse... 

Smoothing, while preserving edges, has always been a major challenge in image processing. In this paper, we propose a new approach that uses segmentation in order to avoid inter-region smoothing thus preserving the edges. It is common to smooth the image prior to region growing. The opposite procedure does not work properly in the presence of noise since region growing is very noise sensitive. To overcome this difficulty we adapted a robust region growing algorithm. Since region growing is very resource consuming, we do not perform it for every pixel. Instead, we divide the image into a number of overlapping blocks for which we carry out the segmentation. Then, we use the results for some of...