A busca por padrões em imagens é um problema clássico em visão computacional e consiste em detectar a presença de uma dada máscara em uma imagem digital. Tal tarefa pode se tornar consideravelmente mais complexa com a invariância aos aspectos da imagem tais como rotação, escala, translação, brilho e contraste (RSTBC - rotation, scale, translation, brightness and contrast). Um algoritmo de busca de máscara foi recentemente proposto. Este algoritmo, chamado de Ciratefi, é invariante aos aspectos RSTBC e mostrou-se bastante robusto. Entretanto, a execução deste algoritmo em um computador convencional requer diversos segundos. Além disso, sua implementação na forma mais geral em hardware é difícil pois há muitos parâmetros ajustáveis. Este trabalho propõe o projeto de um software que gera automaticamente módulos compiláveis em Hardware Description Logic (VHDL) que implementam o filtro circular do algoritmo Ciratefi em dispositivos Field Programmable Gate Array (FPGA). A solução proposta acelera o tempo de processamento de 7s (em um PC de 3GHz) para 1,367ms (em um dispositivo Stratix III da Altera). Esta performance excelente (mais do que o necessário em sistemas em tempo-real) pode levar a sistemas de visão computacional de alta performance e de baixo custo. / Template matching is a classical problem in computer vision. It consists in detecting the presence of a given template in a digital image. This task becomes considerably more complex with the invariance to rotation, scale, translation, brightness and contrast (RSTBC). A novel RSTBC-invariant robust template matching algorithm named Ciratefi was recently proposed. However, its execution in a conventional computer takes several seconds. Moreover, the implementation of its general version in hardware is difficult, because there are many adjustable parameters. This work proposes a software that automatically generates compilable Hardware Description Logic (VHDL) modules that implement the circular filter of the Ciratefi template matching algorithm in Field Programmable Gate Array (FPGA) devices. The proposed solution accelerates the time to process a frame from 7s (in a 3GHz PC) to 1.367ms (in Altera Stratix III device). This excellent performance (more than the required for a real-time system) may lead to cost-effective high-performance coprocessing computer vision systems.

Für den zuverlässigen Betrieb von Hochspannungsfreileistungen ist es notwendig, die an ihnen eingesetzten Isolatoren regelmäßig zu überprüfen. Somit können gravierendere Beschädigungen vorgebeugt werden. Für diese Überprüfung sind verschiedene Verfahren geeignet. Die Brandmalerkennung ist dabei noch ein relativ neues Verfahren. Mit ihrer Hilfe ist es jedoch möglich auch kleinste Beschädigungen zu erkennen. Im Rahmen dieser Bachelorarbeit wird ein neues Verfahren zur Erkennung von Brandmalen vorgestellt. Dieses verwendet einen Template-Matching-Algorithmus zum Finden der Isolatoren. Dessen Erkennungsrate liegt bei 90,18 %. Alle damit gefundenen Isolatoren untersucht man nach Brandmalen. Diese werden segmentiert und durch ein Connected-Component-Labeling lokalisiert. Insgesamt konnten 71,05% der Brandmale erkannt werden. So wurde der Zustand von 88,19% der Isolatoren korrekt bestimmt.

Human face detection and recognition techniques play an important role in applica- tions like face recognition, video surveillance, human computer interface and face image databases. Using color information in images is one of the various possible techniques used for face detection. The novel technique used in this project was the combination of various techniques such as skin color detection, template matching, gradient face de- tection to achieve high accuracy of face detection in frontal faces. The objective in this work was to determine the best rotation angle to achieve optimal detection. Also eye and mouse template matching have been put to test for feature detection.

Respiration introduces large tumor motion in the thoracic region which influences treatment outcome for lung cancer patients. Tumor motion management techniques require characterization of temporal tumor motions because tumor motion varies patient to patient, day to day and cycle to cycle. This work develops a markerless algorithm to estimate 3 dimensional (3D) lung-tumor trajectories on free breathing cone beam computed tomography (CBCT) projections, which are 2 dimensional (2D) sequential images rotating about an axis and are used to reconstruct 3D CBCT images. A gold standard tumor trajectory is required to guide the algorithm development and estimate the tumor detection accuracy for markerless tracking algorithms. However, a sufficient strategy to validate markerless tracking algorithms is lacking. A validation framework is developed based on fiducial markers. Markers are segmented and marker trajectories are xiv obtained. The displacement of the tumor to the marker is calculated and added to the segmented marker trajectory to generate reference tumor trajectory. Markerless tumor trajectory estimation (MLTM) algorithm is developed and improved to acquire tumor trajectory with clinical acceptable accuracy for locally advanced lung tumors. The development is separate into two parts. The first part considers none tumor deformation. It investigates shape and appearance of the template, moreover, a constraint method is introduced to narrow down the template matching searching region for more precise matching results. The second part is to accommodate tumor deformation near the end of the treatment. The accuracy of MLTM is calculated and compared against 4D CBCT, which is the current standard of care. In summary, a validation framework based on fiducial markers is successfully built. MLTM is successfully developed with or without the consideration of tumor deformation with promising accuracy. MLTM outperforms 4D CBCT in temporal tumor trajectory estimation.

Human face detection and recognition techniques play an important role in applica- tions like face recognition, video surveillance, human computer interface and face image databases. Using color information in images is one of the various possible techniques used for face detection. The novel technique used in this project was the combination of various techniques such as skin color detection, template matching, gradient face de- tection to achieve high accuracy of face detection in frontal faces. The objective in this work was to determine the best rotation angle to achieve optimal detection. Also eye and mouse template matching have been put to test for feature detection.

No description available.

No description available.

No description available.

Made available in DSpace on 2016-08-17T14:53:12Z (GMT). No. of bitstreams: 1 Irlandino Almeida.pdf: 795180 bytes, checksum: 9f8896e26b58665bad7d75b665940404 (MD5) Previous issue date: 2007-09-17 / We present a semi-automatic methodology for three-dimensional reconstruction using an image-based modeling environment. This is done through an initial camera calibration step that makes possible to the viewer, identify objects in the 2D acquired image and them get its position from the camera parameters. This methodology can be applied in the real word model reconstruction like electric and industry installations. / Apresentamos uma metodologia semi-automática para reconstrução tridimensional utilizando um ambiente de modelagem no qual o usuário, a partir de uma determinada imagem e de pontos nela selecionados com seu correspondente no espaço do mundo, realiza a calibração da câmera. Após este processo, a cena poderá ser composta a partir de objetos pré-estabelecidos fornecidos pelo ambiente de modelagem tridimensional baseado na imagem. Esta metodologia pode ser aplicada na reconstrução de modelos do mundo real, como instalações elétricas e industriais.

In the scheme of Non Destructive Testing (NDT), defect detection is an important process. Traditional image processing techniques have successfully been used for defect recognition. Usage of machine learning techniques is still in the initial stages of development. Convolution Neural Networks (CNN) is widely used for object classification one such scenario is defect classification in weld tubes. With the advent of deep learning techniques such as transfer learning, we can transfer knowledge gained in one domain successfully into other. Pre-trained models successfully learn features from large scale datasets that can be used for in domains having sparse data and smaller datasets. The aim of this work is to help a manual inspector in recognition of defects on the weld tubes. With a given set of images, we proceed by forming unique pipeline architecture for automatic defect recognition. The research in this thesis focuses on extraction of welds using image segmentation techniques, creating a dataset of defects and using it to on pre-trained Convolution Neural Networks of VGG16, VGG19, Inception V3 and ResNet101. We evaluate the models on different metrics finding the best suited model for the created dataset. Further a prototype sliding window solution is used to find defects over the extracted weld region. We also present the limitations of this approach and suggest modifications that can be implemented in the future.