1,721,204 research outputs found
Deep Generative Models for Realistic Image Anonymization
No full textThe following pages explore the use of generative models for realistic image anonymization. In summary, this thesis aims to address two primary objectives. First, develop generative models for synthesizing human figures for the purpose of anonymization. Secondly, evaluate the impact of anonymization on the development of computer vision algorithms.
This thesis culminates into four key contributions. First, it introduces Deep Privacy, an open-source framework for realistic anonymization of human faces and bodies. Deep Privacy is the first framework to effectively handle the challenges of in-the-wild image anonymization, such as handling overlapping objects, partial bodies, and extreme poses. Secondly, a variety of Generative Adversarial Networks (GANs) are proposed for synthesizing realistic human figures. To the best of our knowledge, the proposed GANs are the first to synthesize human figures in-the-wild effectively. The third contribution comprises of two open-source datasets, namely Flickr Diverse Faces (FDF) and Flickr Diverse Humans (FDH). Unlike previous datasets, FDF and FDH are large-scale and diverse datasets consisting of unfiltered images that capture the complexities of realistic image anonymization. Finally, the thesis presents an empirical evaluation of Deep Privacy and compare it to traditional anonymization. Specifically, the impact of anonymization is evaluated for training computer vision models, with a focus on autonomous vehicle settings.
This thesis demonstrates that realistic anonymization is a superior alternative to traditional methods and a promising method to replace privacy-sensitive data with artificial data. We are confident that our open-source framework and datasets will be highly useful for practitioners and researchers seeking to anonymize their visual data
Applications of Computer Vision and Deep Learning for Digital Rock Analysis
No full textDigital rock analysis utilizes various tools available in computer science to represent and process rock data in a way that enhances our understanding of its geological properties. It has become an essential tool for understanding rock samples’ physical and chemical properties, which is crucial for exploration and production. It is a multi-step process that includes image acquisition, registration, superresolution and segmentation. This thesis proposes various methods to improve the individual steps in this process to improve the overall digital rock analysis workflow.
One of the most critical steps in digital rock analysis is obtaining representative and high-quality rock images. This step is necessary for high accuracy in downstream tasks such as fluid flow simulations. The imaging of rock samples is done via micro-computed tomography (micro-CT) or electron microscopy. The resolution of the images obtained from micro-CT scanning can often be limited for a specific task requiring electron microscopes. However, using an electron microscope presents its own challenges, such as high cost and limited field of view.
First of all, this thesis addresses the limitations in the image acquisition process with the help of upsampling the low-resolution rock images. This process is called image super-resolution. Two deep learning-based super-resolution methods have been presented in the first two papers in this work that can potentially improve the digital rock workflow by enhancing image quality.
A critical and time-consuming aspect of digital rock analysis is image registration, which aligns multiple images of the same rock sample. Without registration, correlating different images of the same rock samples is impossible. The algorithm developed in the third paper, a rigid 3D-3D registration algorithm, is a tool that can finish a registration job in seconds instead of hours taken by current industrial image registration tools.
A rock sample can contain multiple mineral types and regions with different prop erties. Accurately identifying those regions is the first step in determining the properties of the individual parts and, eventually, the whole sample. Rock typing is the process of identifying those regions. It is an essential step in digital rock workflow, commonly performed manually. The second last paper in this thesis presents a deep learning-based method that takes the first steps towards improving the rock typing of laminar rocks.
Scanning the rock samples is costly and time-consuming. Therefore, it is attractive to use deep generative models to generate a representative sample of digital rocks that can be utilized in the workflow. In the fifth and final paper, this thesis presents a novel Diffusion model-based 2D to 3D image generation method. Using the proposed method, a complete 3D image of a rock can be generated using only a single 2D slice, thus addressing the scarcity of 3D data.
In summary, the contributions of this thesis improve the various steps involved in the digital rock analysis workflow using deep learning and conventional computer vision-based methods
SmartRocks: Artificial Intelligence Applications in Digital Rock Physics
No full textThe characterization of subsurface rock’s physical properties plays a crucial role in multiple fields, including geophysics, petroleum, carbon capture & storage, and water resource management. In recent years, Digital Rock Analysis (DRA) has emerged as an effective method for rock characterization, offering cost-effective, non-destructive, and digital assessments of rock samples.
This thesis delivers an in-depth study of artificial intelligence (AI) applications for the field of DRA. The work presented here aims to advance image analysis and generation techniques with the potential to revolutionize current methods and significantly enhance the analysis capabilities of DRA.
The thesis covers multiple projects and four publications that address key problems within the field of DRA. These problems revolve around achieving accurate segmentation, enhancing image resolution, conducting precise image registration, and generating 3D microstructures from 2D images. To tackle these problems, advanced deep learning techniques such as generative adversarial networks (GANs), transformers, and denoising diffusion probabilistic models (DPM) have been employed.
With a strong focus on industry applications, the methodologies presented in this thesis emphasize generalization, scalability, and robustness. This emphasis is particularly crucial when dealing with large-scale 3D image data in diverse digital rock analysis (DRA) scenarios. The research outcomes have been successfully integrated into the SmartRock platform, a web-based service platform, thereby enhancing accessibility for researchers and engineers to utilize the models and tools proposed in this thesis.
By integrating cutting-edge AI technologies into the DRA industry, this research aims to improve DRA workflows, streamline processes, and promote the overall progress of the field, leading to greater efficiency, cost reduction, and enhanced characterization accuracy
Going Beyond Counting First Authors in Author Co-citation Analysis
Full text linkThe present study examines one of the fundamental aspects of author co-citation analysis (ACA) - the way co-citation
counts are defined. Co-citation counting provides the data on which all subsequent statistical analyses and mappings
are based, and we compare ACA results based on two different types of co-citation counting - the traditional type that
only counts the first one among a cited work's authors on the one hand and a non-traditional type that takes into
account the first 5 authors of a cited work on the other hand. Results indicate that the picture produced through this non-traditional author co-citation counting contains more coherent author groups and is therefore considerably clearer. However, this picture represents fewer specialties in the research field being studied than that produced through the traditional first-author co-citation counting when the same number of top-ranked authors is selected and analyzed. Reasons for these effects are discussed
Ultrasound in image-guided spine surgery: Enabling technologies and first steps
Full text linkMost people experience back pain at some point in their life. While most of these conditions are treated nonoperatively, surgical treatment has been shown to be both effective and cost effective compared to nonoperative care for both intervertebral disc herniation and spinal stenosis in selected patients. Surgical navigation systems enabling image guidance based on preoperative or intraoperative computed tomography (CT) images have found some use in spine surgery. Here, they are most frequently used for spinal fusion, and the benefits of image guidance in such procedures, under given conditions, have been documented in several studies. In spite of this, few spine surgeons use navigation routinely.
High cost is one of the most important barriers to a more widespread adoption of navigation systems in spine surgery. Extending the use of navigation to more than just fusion procedures, and to soft tissue procedures such as disk herniations in particular, could help the surgeon substantiate the cost of the equipment. A big step in this direction would be to enable navigation based on other imaging modalities than CT, such as magnetic resonance (MR) or ultrasound imaging. In this work, the goal has therefore been to enable the use of ultrasound imaging both for intraoperative imaging and for registration of preoperative MR images to the patient.
Many spine procedures are today performed with a microsurgical approach, and the small incisions used in these procedures prohibit the use of standard probes. Our group has therefore, in a previous project, developed a new probe specifically designed to enable ultrasound imaging through such small incisions. The main part of the work has been directed towards enabling tracking and navigation with this probe. In addition, we have studied methods for registering MR images to ultrasound images of the spine.
In Paper A, we looked at methods for reconstructing three-dimensional image volumes from tracked two-dimensional ultrasound images. Both different means of capturing the original ultrasound images and different reconstruction algorithms were thoroughly compared. We found that the differences were small, and while the various methods showed different strengths and weaknesses, the overall result was that they could not be separated.
In Paper B, we explored the feasibility of using electromagnetic (EM) tracking in an operating room setting, both alone and in combination with a robotic C-arm. We also compared the performance of the standard EM field generator with a new prototype designed specifically for use with fluoroscopic imaging equipment. We found that while the accuracy decreased considerably with the C-arm inside the operating field, the measurements were still stable. We thus concluded that by implementing a suitable static correction scheme, the tracking system and the C-arm could potentially be used together.
In Paper C, we presented a new method for ultrasound probe calibration, which is the process of finding the spatial relationship between the coordinate system of the tracking sensor that is integrated in the ultrasound probe and the coordinate system of the ultrasound images generated by the probe. In a research setting, such as ours, new probes are tested regularly, and the method was therefore designed to be used with a large variety of probes without any adaption. The method was tested on three very different probes demonstrating both great versatility and high accuracy.
In Paper D, we developed a method for registration of preoperative MR images to the patient by means of intraoperative ultrasound imaging using a tracked ultrasound probe. The method segmented the posterior bone surface from both the ultrasound images and the MR images and registered the two surfaces to each other using a modified
version of the Iterative Closest Point algorithm. For this paper, the method was only tested on one subject, but the accuracy of the registration on this subject was clinically relevant, and we concluded that the method was promising.
In conclusion, we have developed and tested technology that enables tracking of small, intraoperative ultrasound probes and allows the generation of three-dimensional volumes suitable for navigation from such images. We have also investigated the use of intraoperative ultrasound imaging for registration of preoperative CT and MR images to the spine. The latter is, however, a work in progress, as the methods that we have tested have so far have not been sufficiently robust for clinical us
Improved Bronchoscopy by new image guided Approach
Full text linkNavigation in bronchoscopy has developed into a feasible approach for lung diagnostics since introduced in 1998. The concept combines computer generated models from patient CT data with position and orientation tracking of the bronchoscope tip and/or other tools. Trials using navigational bronchoscopy have demonstrated increased diagnostics success rates. Despite higher diagnostic success rates and two decades of development in navigation for bronchoscopy, the use of the technology is still not common in lung diagnostics.
The pulmonologist is trained surveying and assessing CT and steering via video display from a flexible scope during bronchoscopy. Expanding the conventional work with additional display through techniques such as navigational bronchoscopy can be expensive to integrate in smaller hospitals and complicated for pulmonologists without proper training. Also, existing VB approaches only offers endoluminal view, an airway segmentation as a “road map” to lesions in lungs, and in best circumstances additional models of segmented lesions and vessels.
New solutions specifically for bronchoscopy visualization application was developed in this project. This thesis presents a development and evaluation of a new visualization approach for planning and guidance in bronchoscopy; Anchored to Centerline Curved Surface (ACCuSurf), consisting of more complete view for navigated bronchoscopy in tube-like structures. The technique may also be combined with other methods such as VB, PET and ultrasound images, by adding these data sources to the display. At the same time as providing overview of the lungs and tools, the ACCuSurf can be zoomed in and show more anatomical details than the conventional endoluminal view. First, a comparison of different approaches to airway segmentation was carried out to establish a route to target. Secondly, the ACCuSurf was developed by slicing the segmented airways in half, creating a 3D volume representing surrounding anatomy along the path to target. Finally, the ACCuSurf method was assessed by pulmonologists using it as a planning tool before performing bronchoscopy on a phantom with a mixed data set from a patient and the phantom. The conventional 2D (axial, sagittal, coronal) visualisation was comparison reference. The study is an effort to ease and simplify visualisation for navigation in bronchoscopy
Variations on the Author
Full text link“Variations on the Author” discusses two of Eduardo Coutinho’s recent films (Um Dia na Vida, from 2010, and Últimas Conversas, posthumously released in 2015) and their contribution to the general question of documentary authorship. The director’s filmography is characterized by a consistent yet self-effacing form of authorial self-inscription: Coutinho often features as an interviewer that rather than express opinions propels discourses; an interviewer that is good at listening. This mode of self-inscription characterizes him as an author who is not expressive but who is nonetheless markedly present on the screen. In Um Dia na Vida, however, Coutinho is completely absent form the image, while Últimas Conversas, on the contrary, includes a confessional prologue that moves the director from the margins to the center of his films. This article examines the ways in which these works stand out in the filmography of a director who offers new insights into the notion of cinematic authorship
Appropriate Similarity Measures for Author Cocitation Analysis
Full text linkWe provide a number of new insights into the methodological discussion about author cocitation analysis. We first argue that the use of the Pearson correlation for measuring the similarity between authors’ cocitation profiles is not very satisfactory. We then discuss what kind of similarity measures may be used as an alternative to the Pearson correlation. We consider three similarity measures in particular. One is the well-known cosine. The other two similarity measures have not been used before in the bibliometric literature. Finally, we show by means of an example that our findings have a high practical relevance.information science;Pearson correlation;cosine;similarity measure;author cocitation analysis
Dispelling the Myths Behind First-author Citation Counts
Full text linkWe conducted a full-scale evaluative citation analysis study of scholars in the XML research field to explore just how different from each other author rankings resulting from different citation counting methods actually are, and to demonstrate the capability of emerging data and tools on the Web in supporting more realistic citation counting methods. Our results contest some common arguments for the continued
use of first-author citation counts in the evaluation of scholars, such as high correlations between author rankings by first-author citation counts and other citation
counting methods, and high costs of using more realistic citation counting methods that are not well-supported by the ISI databases. It is argued that increasingly available digital full text research papers make it possible for citation analysis studies to go beyond what the ISI databases have directly supported and to employ more
sophisticated methods
- …
