• 2020-2024close
• Open Access close
• Closed Access close
• Restricted close
• Publications close
• Research data close
• Research software close
• Other research products close
• CZ close
• Digitální knihovna VUT close
• North American Studies close

This work is focused on the image processing using the OpenCV library and detectingmoving objects in video using convolutional neural networks. The created application can detectmoving objects in the video and contains additional functionality. The application includes theYOLO convolutional model, which helps to detect moving objects.

Phoebe Apperson Hearst měla velmi úspěšného vlastního syna Williama, který byl ale více jako jeho otec: tvrdý obchodník. Našla však jemnou, uměleckou duši v malíři Orrinu Peckovi (1860–1921), který byl údajně gay a který ji, ještě za života své vlastní matky, začal oslovovat „má druhá mámo.“ Na základě podrobného výzkumu jejich vzájemné korespondence v Peckově pozůstalosti se můžeme ptát, jak moc si byla progresivní, bohatá žena 19. století, jakou byla Phoebe Hearst, vědoma Peckovy sexuality a pokud ano, jestli s tím neměla problém, nebo šlo o nevyřčené tajemství mezi nimi? Jejich příběh představí historik umění Ladislav Zikmund-Lender. Phoebe Apperson Hearst had a very successful son of William, but he was more like his father: a tough businessman. However, she found a delicate, artistic soul in the painter Orrin Peck (1860–1921), who was allegedly gay and who, while still his own mother's life, began to address her as “my second mother.” Based on a detailed study of their correspondence in Peck's estate, we may ask how much a progressive, rich 19th-century woman like Phoebe Hearst was aware of Peck's sexuality, and if so, if she had no problem with it, or was it an unspoken secret between them? Their story will be presented by art historian Ladislav Zikmund-Lender.

In this paper, we present two pipelines in order to reduce the feature space for anomalydetection using the One Class SVM. As a first stage of both pipelines, we compare the performanceof three convolutional autoencoders. We use the PCA method together with t-SNE as the first pipelineand the reconstruction errors based method as the second. Both methods have potential for theanomaly detection, but the reconstruction error metrics prove to be more robust for this task. Weshow that the convolutional autoencoder architecture doesn’t have a significant effect for this task andwe prove the potential of our approach on the real world dataset.

We propose the model combining convolutional neural network with multiple instancelearning in order to localize the premature atrial contraction and premature ventricular contraction.The model is based on ResNet architecture modified for 1D signal processing. Model was trainedon China Physiological Signal Challenge 2018 database extended by manually labeled ground truthpositions of premature complexes. The presented method did not reach satisfying results in PAClocalization (with dice = 0.127 for avg-pooling implementation). On the other hand, results of localizationof PVCs were comparable with other published studies (with dice = 0.952 for avg-poolingimplementation).

This paper presents an evaluation of deep learning architectures designed for modulationrecognition. The evaluation inspects, whether the architectures behave in the same way as they didon the dataset they were designed on. The architectures are trained and tested on two different radiomodulation datasets. This results in proposing additional binary classification as a method to reducemisclassification of QAM modulation types in one of the datasets.

In 1939, due to WWII and the Nuremberg Laws, the revolutionary Czech structural engineer Jaroslav J. Polívka arrived in the United States. After his arrival, he started a research job at UC Berkeley, renewed his engineering practice, and offered his services to the US military as many businesspersons did during this era. Polívka worked for Henry Kaiser who turned Richmond, CA, into a vibrant, fast developing workers city. New residential districts, hospitals, hangars, docks, and warehouses were built there. Henry Kaiser approached the structural development of the city in the same way he revolutionized the construction of battleships: from prefabricated, standardized parts. He supported research and development of new technologies. Mobile, round-shaped hospitals from prefabricated aluminum frames were one of the results of that research. In 1946, Jaroslav J. Polívka introduced himself to the “starchitect” Frank Lloyd Wright. A productive mutual co-operation that lasted 13 years and resulted in eight spectacular projects had started and Polívka, who had been working on extensive research both at UC Berkeley and Stanford University, came up with many technological, structural, and material innovations over the period. In 1957, Henry Kaiser funded the construction of one of the two geodesic domes designed by Richard Buckminster Fuller in Hawaii and in the process he invited Frank Lloyd Wright to consult the project. Jaroslav J. Polívka was probably in direct contact with Fuller, since he wanted to include him in his unfinished project of an encyclopedia of the world-famous structural engineers. On this particular story and a social matrix evolving around Henry Kaiser and Frank Lloyd Wright, the lecture seeks to rethink architectural global modernism as a cooperative project rather than a series of individual innovations manifested by isolated genius figures.

Automatic segmentation of the biological structures in micro-CT data is still a challengesince the object of interest (craniofacial cartilage in our case) is commonly not characterized by uniquevoxel intensity or sharp borders. In recent years, convolutional neural networks (CNNs) have becomeexceedingly popular in many areas of computer vision. Specifically, for biomedical image segmentationproblems, U-Net architecture is widely used. However, in case of micro-CT data, there isa question whether 3D CNN would not be more beneficial. This paper introduces CNN architecturebased on V-Net as well as the methodology for data preprocessing and postprocessing. The baselinearchitecture was further optimized using advanced techniques such as Atrous Spatial Pyramid Pooling(ASPP) module, Scaled Exponential Linear Unit (SELU) activation function, multi-output supervisionand Dense blocks. For network learning, modern approaches were used including learning ratewarmup or AdamW optimizer. Even though the 3D CNN do not outperform U-Net regarding the craniofacialcartilage segmentation, the optimization raises the median of Dice coefficient from 69.74 %to 80.01 %. Therefore, utilizing these advanced techniques is highly encouraged as they can be easilyadded to any U-Net-like architecture and may remarkably improve the results.

L'extraction des empreintes des bâtiments à partir de vastes données de télédétection à très haute résolution spatiale (VHSR) est cruciale pour diverses applications, notamment l'arpentage, les études urbaines, l'estimation de la population, l'identification des établissements informels et la gestion des catastrophes. Bien que les réseaux de neurones convolutifs (CNN) soient couramment utilisés à cette fin, leur efficacité est limitée par les limites de la capture des relations à long terme et des détails contextuels en raison de la nature localisée des opérations de convolution. Cette étude présente le transformateur de masque à attention masquée (Mask2Former), basé sur le transformateur Swin, pour l'extraction de l'empreinte du bâtiment à partir d'images satellitaires à grande échelle. Pour améliorer la capture d'informations sémantiques à grande échelle et extraire des caractéristiques multi-échelles, un transformateur de vision hiérarchique à fenêtres décalées (Swin Transformer) sert de réseau fédérateur. Une analyse approfondie compare l'efficacité et la généralisabilité de Mask2Former avec quatre modèles CNN (PSPNet, DeepLabV3+, UpperNet-ConvNext et SegNeXt) et deux modèles basés sur des transformateurs (UpperNet-Swin et SegFormer) présentant différentes complexités. Les résultats révèlent des performances supérieures des modèles basés sur les transformateurs par rapport à leurs homologues basés sur CNN, mettant en évidence une généralisation exceptionnelle dans diverses zones de test avec des structures, des hauteurs et des tailles de bâtiment variables. Plus précisément, Mask2Former avec l'épine dorsale du transformateur Swin atteint une intersection moyenne sur l'union entre 88 % et 93 %, ainsi qu'un score F moyen (score mF) allant de 91 % à 96,35 % dans divers paysages urbains. La extracción de huellas de edificios a partir de datos extensos de teledetección de muy alta resolución espacial (VHSR) es crucial para diversas aplicaciones, incluidas la topografía, los estudios urbanos, la estimación de la población, la identificación de asentamientos informales y la gestión de desastres. Aunque las redes neuronales convolucionales (CNN) se utilizan comúnmente para este propósito, su efectividad se ve limitada por las limitaciones en la captura de relaciones de largo alcance y detalles contextuales debido a la naturaleza localizada de las operaciones de convolución. Este estudio presenta el transformador de máscara de atención enmascarada (Mask2Former), basado en el transformador Swin, para la extracción de huellas de edificios a partir de imágenes satelitales a gran escala. Para mejorar la captura de información semántica a gran escala y extraer características multiescala, un transformador de visión jerárquico con ventanas desplazadas (Swin Transformer) sirve como red troncal. Un extenso análisis compara la eficiencia y la generalización de Mask2Former con cuatro modelos CNN (PSPNet, DeepLabV3+, UpperNet-ConvNext y SegNeXt) y dos modelos basados en transformadores (UpperNet-Swin y SegFormer) con diferentes complejidades. Los resultados revelan un rendimiento superior de los modelos basados en transformadores sobre sus homólogos basados en CNN, mostrando una generalización excepcional en diversas áreas de prueba con diferentes estructuras de edificios, alturas y tamaños. Específicamente, Mask2Former con la columna vertebral del transformador Swin logra una intersección media sobre la unión entre el 88% y el 93%, junto con una puntuación F media (puntuación mF) que oscila entre el 91% y el 96,35% en varios paisajes urbanos. Extracting building footprints from extensive very-high spatial resolution (VHSR) remote sensing data is crucial for diverse applications, including surveying, urban studies, population estimation, identification of informal settlements, and disaster management. Although convolutional neural networks (CNNs) are commonly utilized for this purpose, their effectiveness is constrained by limitations in capturing long-range relationships and contextual details due to the localized nature of convolution operations. This study introduces the masked-attention mask transformer (Mask2Former), based on the Swin Transformer, for building footprint extraction from large-scale satellite imagery. To enhance the capture of large-scale semantic information and extract multiscale features, a hierarchical vision transformer with shifted windows (Swin Transformer) serves as the backbone network. An extensive analysis compares the efficiency and generalizability of Mask2Former with four CNN models (PSPNet, DeepLabV3+, UpperNet-ConvNext, and SegNeXt) and two transformer-based models (UpperNet-Swin and SegFormer) featuring different complexities. Results reveal superior performance of transformer-based models over CNN-based counterparts, showcasing exceptional generalization across diverse testing areas with varying building structures, heights, and sizes. Specifically, Mask2Former with the Swin transformer backbone achieves a mean intersection over union between 88% and 93%, along with a mean F-score (mF-score) ranging from 91% to 96.35% across various urban landscapes. يعد استخراج آثار المباني من بيانات الاستشعار عن بعد عالية الدقة المكانية (VHSR) أمرًا بالغ الأهمية للتطبيقات المتنوعة، بما في ذلك المسح والدراسات الحضرية وتقدير السكان وتحديد المستوطنات غير الرسمية وإدارة الكوارث. على الرغم من استخدام الشبكات العصبية الالتفافية (CNNs) بشكل شائع لهذا الغرض، إلا أن فعاليتها مقيدة بالقيود المفروضة على التقاط العلاقات طويلة المدى والتفاصيل السياقية بسبب الطبيعة المحلية لعمليات الالتفاف. تقدم هذه الدراسة محول قناع الانتباه المقنع (Mask2Former)، استنادًا إلى محول Swin، لبناء استخراج بصمة من صور الأقمار الصناعية واسعة النطاق. لتعزيز التقاط المعلومات الدلالية واسعة النطاق واستخراج الميزات متعددة المقاييس، يعمل محول الرؤية الهرمي المزود بنوافذ متحركة (محول Swin) كشبكة أساسية. يقارن التحليل الشامل كفاءة وتعميم Mask2Former مع أربعة نماذج CNN (PSPNet و DeepLabV3 + و UpperNet - ConvNext و SegNeXt) ونموذجين قائمين على المحولات (UpperNet - Swin و SegFormer) يتميزان بتعقيدات مختلفة. تكشف النتائج عن أداء متفوق للنماذج القائمة على المحولات على نظيراتها القائمة على CNN، مما يعرض تعميمًا استثنائيًا عبر مناطق اختبار متنوعة ذات هياكل بناء وارتفاعات وأحجام مختلفة. على وجه التحديد، يحقق Mask2Former مع العمود الفقري لمحول Swin تقاطعًا متوسطًا بين 88 ٪ و 93 ٪، إلى جانب متوسط درجة F (درجة mF) تتراوح من 91 ٪ إلى 96.35 ٪ عبر مختلف المناظر الطبيعية الحضرية.

A convolutional neural network is often mentioned as one of the deep learning methods that requires a large amount of training data. Questioning this belief, this paper explores the applicability of classification based on a shallow net structure trained on a small data set in the~context of patient posture classification based on data from a pressure mattress. Designing a CNN often presents a complex problem, especially without a universally applicable approach, allowing many diverse structural possibilities and training settings. We tested various training options and layer configurations to provide an overview of influential parameters for posture classification. Experiments show encouraging results with the leave-one-out cross-validation accuracy of 93.1% of one of the evaluated CNN structures and its hyperparameter settings.