2023

Permanent URI for this collectionhttps://hdl.handle.net/20.500.14570/3921

Browse

Search Results

Now showing 1 - 10 of 13
  • Thumbnail Image
    Item
    Weakly-supervised tumor segmentation in computed tomography scans
    (2023) Zakharchenko, Iryna
    This research focuses on the topic of weakly-supervised tumor segmentation. The proposed pipeline involves the usage of a classification model to make predictions regarding the presence of a tumor in an image. Subsequently, the CAM (Class Activation Mapping) approach is employed to identify the most relevant regions within the image as determined by the model. The underlying concept is that the model will learn to identify tumor regions, resulting in higher activations in those areas. The advantage of the weakly supervised approach is its ability to learn from a smaller dataset, requiring only image-level labels in our case. By implementing the proposed pipeline, specifically using the Score-CAM technique.
  • Thumbnail Image
    Item
    Monitoring Sea Water Pollution from Satellite data
    (2023) Stepanov, Yevhen
    This study is motivated by the Ocean Decade declared by the United Nations. We employed machine learning techniques to detect and delineate areas of pollution in the coastal zone of Great Britain, utilizing pollution reports from the Department for Environment Food And Rural Affairs (DEFRA) and the ocean monitoring datasets from the European Space Agency (ESA). In this study, feature engineering was performed on chlorophyll concentration data. Two datasets were constructed: one with statistical metrics (mean, median, standard deviation, and percentiles) as features, and another with individual cells of the chlorophyll concentration matrix as features, utilizing different matrix n sizes, where n is from 3 to 11, where each element or pixel of the matrix represented a 1km × 1km area. Logistic regression, decision trees, random forest classifier, gradient boosting classifier, and LeNet models were applied. Hyper parameter tuning was conducted to optimize the performance of each model. Among the models, the gradient boosting classifier achieved the highest accuracy of 95.21%. Additionally, the F1 score was determined to be 0.2445, the ROC AUC was 0.7659, and the precision-recall AUC (PR-AUC) was found to be 0.1821. Detecting and delineating areas of pollution can greatly assist cleaning services in efficiently carrying out their job, resulting in improved remediation and restoration efforts. The identification of pollution areas holds significant implications for the fishing industry, as it enables informed decision-making regarding fishing practices and resource management, ensuring the sustainability and viability of the sector. Moreover, the accurate detection and delineation of pollution areas have the potential to generate substantial economic, social, and environmental benefits by facilitating targeted interventions, protecting ecosystems, preserving marine resources, and fostering a healthier and more resilient environment. The findings of this study provide valuable insights into the efficacy of classification approaches in identifying and mapping pollution sites in coastal regions using pollution reports from DEFRA.
  • Thumbnail Image
    Item
    Corner localization and camera calibration from imaged lattices
    (2023) Stadnik, Andrii
    This thesis proposes a model-based approach to improve the detection of calibration board fiducials from calibration imagery taken by wide-angle or fisheye lenses. From a single image, we estimate the camera model and project the calibration board into the image to guide the search for missed detections and reject spurious detections. In addition, we propose a classifier to label ambiguous detections that are geometrically plausible given the estimated camera model and imaged board. The proposed method addresses shortcomings of the state-of-the-art, which struggle to reliably detect board fiducials at the extents of the image, where the lens distortion is most observable. The proposed method recovers additional corners that can be used to place additional constraints on the non-convex camera calibration problem, which improves the likelihood of convergence to a global minimum. The code for this paper is available on GitHub.
  • Thumbnail Image
    Item
    Ocean surface visibility prediction
    (2023) Prypeshniuk, Volodymyr
    Seawater transparency is an indispensable ecological parameter with substantial impacts on the health and productivity of aquatic ecosystems. Its significance spans across various industries, including environment protection, fishing and tourism. The fluctuating nature of aquatic systems and their intricate interplay with human activities often induce substantial variability in seawater transparency. This underlines the pressing necessity for effective predictive tools in the stewardship and preservation of our invaluable water resources. Despite the clear importance of water transparency, ocean forecasting remains a considerably understudied field, some work has been done on using satellite for monitoring, but literature is scarce for forecasting with only few simple models explored. There is an evident gap in research and tools focused on predicting changes in this crucial ecosystem, underlining the novelty and urgency of our work. In this research, we aim is to develop a forecasting model that not only excels in precision and speed, but is also flexible enough to encompass a vast array of potential future scenarios. We primarily employed SimVP, a spatio-temporal convolutional neural network, for ocean forecasting purposes. This model was trained using the earth observation data from the Copernicus Marine Service. This data were collected for 20 years of daily observation of water transparency in the marine environment surrounding the UK, with a spatial resolution of 4km x 4km. Our findings showed that SimVP substantially outperformed the baseline models (AutoRegressive Integrated Moving Average (ARIMA) and Simple Exponential Smoothing (SES)) in predicting the next day seawater transparency, demonstrating an improvement of 17.4%, and a notable reduction in the Root Mean Square Error (RMSE) from 2.63 to 2.24, and improvement in inference time efficiency in 66.3 times (334.6 -> 5.04 seconds). We show that this method better performs better on regions with minor variation like Irish Sea or English Channel, and performs worse on regions with high variations like Atlantic Ocean or North Sea. Our study demonstrates the advantage of adopting the spatio-temporal neural network architectures for ocean monitoring and paves the way for future research in adopting advanced machine learning techniques in this field.
  • Thumbnail Image
    Item
    Advancing medical image segmentation via pseudo-labeling of public datasets
    (2023) Mishchenko, Roman
    Our study explores the difficulties and possible resolutions in the domain of medical image segmentation, with a special emphasis on utilizing unlabeled public datasets to improve tumor segmentation. We suggest a strategy that incorporates pseudolabeling methodologies with real-world data to enhance the learning potential of segmentation models. Yet, the findings imply that while improvements in model performance exist, they are not substantial. The research underscores the paramount importance of data quality over quantity, emphasizing that image characteristics influence the effectiveness of the process more than the total number of images.
  • Thumbnail Image
    Item
    Astronomical Data Features Extraction and Citation Prediction
    (2023) Kutsuruk, Vladyslav
    Natural Language Processing methods present promising opportunities for analyzing astronomical data, enabling the extraction of essential information from vast amounts of observations. Yet, applying these techniques to astronomical data presents notable challenges, including the difficulty of astronomical terminology and the diverse range of data sources. In this research, we leverage multiple Natural Language Processing techniques to extract information from astronomical observations with a specific focus on predicting the future citation rate of astronomical telegrams. To achieve this, we create a comprehensive dataset gathering astronomical messages from various sources and utilize techniques such as Named Entity Recognition, doc2vec, word2vec, and topic extraction. Along with this, we enhance the extracted information by incorporating manually created features that capture the characteristics of astronomical telegrams beyond their direct context. These features aim to provide a comprehensive representation of the messages. We then use all the extracted information to predict the future impact of the telegrams, as indicated by their citation counts, using multiple Machine Learning techniques.
  • Thumbnail Image
    Item
    Text-Guided 3D Synthesis with Latent Diffusion Models
    (2023) Kovalenko, Danylo
    The emergence of diffusion models has greatly impacted the field of deep generative models, establishing them as a powerful family of models with unparalleled performance in various applications such as text-to-image, image-to-image, and text-toaudio tasks. In this work, we aim to propose a solution for text-guided 3D synthesis using denoising diffusion probabilistic models, while minimizing the memory and computational requirements. Our goal is to achieve high-quality and high-fidelity 3D object generation conditioned by text or a label in a number of seconds. We propose to use a triplane space parametrization in combination with a Latent Diffusion Model (LDM) to generate smooth and coherent geometry. The LDM is trained on the large-scale text-3d dataset and is used as a latent triplane texture generator. By using a triplane space parametrization, we aim to improve the efficiency of the space representation and reduce the computational cost of synthesis. We will also give a theoretical justification that this kind of parametrization of 3d space is capable of containing not only information about the geometry but also about the color and reflectivity of the figure. Additionally, we use an implicit neural renderer to decode geometry details from triplane textures.
  • Thumbnail Image
    Item
    Text generation with control conditions compliance
    (2023) Konopatska, Oleksandra
    Controllable text generation has emerged as a significant research area, allowing the production of text with desired characteristics. In this work, we investigate the controllability of text generation, exploring the challenges of controlling various aspects of generated text, such as length, parts-of-speech (POS) structure, sentiment, and tense; in addition, we extend our analysis to the task of multi-conditional text generation, which entails the possibility of simultaneous control of several parameters of the generated text. Our research is mainly based on fine-tuned GPT-2, an autoregressive transformerbased model. Using fine-tuned GPT-2, we managed to achieve notable progress in controlling the above-mentioned text attributes; we also present the results of experiments using other approaches, such as diffusion models and ChatGPT. The models are trained on our own dataset, meticulously curated in-house; the evaluation of the generation results is carried out using a comprehensive set of control, fluency, distinctiveness, and repetition metrics. Through rigorous analysis, we assess the performance of studied models in terms of controllability. Length control, in particular, proved to be a challenging aspect, even when employing the largest available models. Nevertheless, our fine-tuned GPT-2 demonstrated promising results, showcasing its capabilities in generating text with desired characteristics. Overall, our findings highlight the possibilities of controllable text generation using fine-tuned GPT-2 and other models. Our work contributes to the ongoing exploration of techniques for improving controllability in text generation. As this field continues to evolve, further research can build upon our analysis and methodologies to enhance controllability and pave the way for more sophisticated text generation systems.
  • Thumbnail Image
    Item
    Split Activation Networks for Neural Fields
    (2023) Kilianovskyi, Mykhailo
    Neural field modeling is a developing area that improves state-of-the-art results in tasks such as 3D scene reconstruction, image manipulation, generative modeling, and other aspects of deep learning. In this work, we present SplitNet, a novel neural network architecture for neural field modeling that combines multiple activation functions in a single layer. We try different techniques to improve performance, such as proper weight initialization, and benchmark its performance on image representation, 3D scene reconstruction, and image classification tasks. As a part of the work, we found a way to improve the performance of previous work on implicit neural networks with sinusoidal activations in a limited setting and study how well this improvement generalizes to other tasks and data.
  • Thumbnail Image
    Item
    Audio spoofing detection
    (2023) Ivashchenko, Dmytro
    Efficient and accurate audio spoofing detection is crucial to ensuring audio-based systems’ security and integrity. Existing methods often mainly focused on the performance of the detection system. This master thesis focuses on the development of advanced techniques that prioritize efficiency while maintaining high detection performance. We introduced the model, consisting of an encoder and a classifier, which can efficiently learn complex representations with a lack of labeled data. We introduce suitable loss functions to effectively distinguish spoofed and bonafide speech in latent space to keep the performance high. The results demonstrate notable improvements in both encoder performance and classification accuracy, highlighting the potential for enhanced self-supervised audio analysis techniques. Keywords: self-supervised learning, audio spoofing detection, automatic speaker verification, audio processing, speech classification