This master thesis studies the learning of dense feature descriptors where camera poses are the only supervisory signal. The use of camera poses as a supervisory signal has only been published once before, and this thesis expands on this previous work by utilizing a couple of different techniques meant increase the robustness of the method, which is particularly important when not having access to ground-truth correspondences. Firstly, an adaptive robust loss is utilized to better differentiate inliers and outliers. Secondly, statistical properties during training are both enforced and adapted to, in an attempt to alleviate problems with uncertainties introduced by not having true correspondences available. These additions are shown to slightly increase performance, and also highlights some key ideas related to prediction certainty and robustness when working with camera poses as a supervisory signal. Finally, possible directions for future work are discussed.

Muitos procedimentos vêm sendo desenvolvidos para auxiliar no diagnóstico precoce do câncer de mama. Devido a subjetividade na interpretação de imagens, os sistemas de diagnóstico auxiliado por computador (CADx) têm oferecido ao especialista uma segunda opinião mais precisa e confiável. Nesse propósito, essa pesquisa apresenta uma metodologia de investigação da potencialidade diagnóstica de um sistema computacional na classificação de achados suspeitos em imagens de ultrassom modo-B e de elastografia da mama. A base de dados foi constituída por 31 lesões malignas e 52 benignas e um conjunto adicional contendo 206 lesões de ultrassom modo-B (144 benignas e 62 malignas) para a realização dos testes de aprendizado de máquina. O contorno foi determinado automaticamente e através do delineamento manual de três radiologistas sob a imagem de ultrassom modo-B e, em seguida, mapeado na imagem elastográfica. As lesões foram classificadas pelo sistema CADx desenvolvido para ultrassom modo-B e elastografia do tipo strain. Os dados foram avaliados por meio da sensibilidade, especificidade e AUC. O sistema CADx desenvolvido proporcionou equivalência diagnóstica para a classificação das lesões a partir das diversas formas de determinação do contorno (manual e automática), permitindo a redução da variabilidade. Além disso, o sistema apontou resultados superiores à análise visual do radiologista que, quando considerado o resultado fornecido pela associação entre as imagens de ultrassom modo-B e elastografia, proporcionou um aumento comparativo de cerca de 7% em sensibilidade e 17,2% em especificidade nos testes com o sistema CADx usando o contorno feito pelo radiologista mais experiente. Além disso, constatou-se uma influência positiva no uso da ferramenta computacional pelos radiologistas, pois, na média, seus índices de sensibilidade e especificidade diagnóstica aumentaram também em relação à situação de análise convencional, passando de 87,1% e 55,8% para 90,3% e 73,1%, respectivamente. / Many procedures have been developed to aid in the early detection and diagnosis of breast cancer. In this context, Computer-Aided Diagnosis (CADx) systems were designed to provide to the specialist a reliable second opinion. This study presents the proposal of investigating the diagnostic ability of a computational system in the characterization of suspicious findings in B-mode ultrasound and breast elastography imaging. The database consisted of 31 malignant and 52 benign lesions and an additional data set containing 206 lesions (144 benign and 62 malignant) seen only on the B-mode ultrasound for performing the machine learning tests. Three radiologists drew manually the contour of the lesions in B-mode ultrasound and we used an automatic technique to segment the lesions. Then, the contour was mapped in the elastography image. The lesions were classified using the CADx system developed for B-mode ultrasound and strain elastography. We calculated the sensitivity, specificity and AUC to evaluate the data. The developed CADx system provided a diagnostic concordance in the classification of breast lesions from the different ways of contour determination (manual and automatic), allowing to reduce the diagnostic variability. In addition, the CADx system showed superior results to the visual analysis of the radiologist. When the radiologist associated both examinations (B-mode ultrasound and elastography), his visual analysis provided 87.10%, 55.77% and 0.714 of sensitivity, specificity and AUC, respectively. When we considered the result provided by the association between B-mode ultrasound and elastography images, the CADx system provided a comparative increase of about 7% of sensitivity and 17.2% of specificity, using the contour delimited by the most experienced radiologist. In addition, a positive influence was observed in the use of the computational tool by radiologists, since, on average, their sensitivity and specificity indexes also increased in relation to the conventional analysis, from 87.1% and 55.8% to 90.3% and 73.1%, respectively.

Muitos procedimentos vêm sendo desenvolvidos para auxiliar no diagnóstico precoce do câncer de mama. Devido a subjetividade na interpretação de imagens, os sistemas de diagnóstico auxiliado por computador (CADx) têm oferecido ao especialista uma segunda opinião mais precisa e confiável. Nesse propósito, essa pesquisa apresenta uma metodologia de investigação da potencialidade diagnóstica de um sistema computacional na classificação de achados suspeitos em imagens de ultrassom modo-B e de elastografia da mama. A base de dados foi constituída por 31 lesões malignas e 52 benignas e um conjunto adicional contendo 206 lesões de ultrassom modo-B (144 benignas e 62 malignas) para a realização dos testes de aprendizado de máquina. O contorno foi determinado automaticamente e através do delineamento manual de três radiologistas sob a imagem de ultrassom modo-B e, em seguida, mapeado na imagem elastográfica. As lesões foram classificadas pelo sistema CADx desenvolvido para ultrassom modo-B e elastografia do tipo strain. Os dados foram avaliados por meio da sensibilidade, especificidade e AUC. O sistema CADx desenvolvido proporcionou equivalência diagnóstica para a classificação das lesões a partir das diversas formas de determinação do contorno (manual e automática), permitindo a redução da variabilidade. Além disso, o sistema apontou resultados superiores à análise visual do radiologista que, quando considerado o resultado fornecido pela associação entre as imagens de ultrassom modo-B e elastografia, proporcionou um aumento comparativo de cerca de 7% em sensibilidade e 17,2% em especificidade nos testes com o sistema CADx usando o contorno feito pelo radiologista mais experiente. Além disso, constatou-se uma influência positiva no uso da ferramenta computacional pelos radiologistas, pois, na média, seus índices de sensibilidade e especificidade diagnóstica aumentaram também em relação à situação de análise convencional, passando de 87,1% e 55,8% para 90,3% e 73,1%, respectivamente. / Many procedures have been developed to aid in the early detection and diagnosis of breast cancer. In this context, Computer-Aided Diagnosis (CADx) systems were designed to provide to the specialist a reliable second opinion. This study presents the proposal of investigating the diagnostic ability of a computational system in the characterization of suspicious findings in B-mode ultrasound and breast elastography imaging. The database consisted of 31 malignant and 52 benign lesions and an additional data set containing 206 lesions (144 benign and 62 malignant) seen only on the B-mode ultrasound for performing the machine learning tests. Three radiologists drew manually the contour of the lesions in B-mode ultrasound and we used an automatic technique to segment the lesions. Then, the contour was mapped in the elastography image. The lesions were classified using the CADx system developed for B-mode ultrasound and strain elastography. We calculated the sensitivity, specificity and AUC to evaluate the data. The developed CADx system provided a diagnostic concordance in the classification of breast lesions from the different ways of contour determination (manual and automatic), allowing to reduce the diagnostic variability. In addition, the CADx system showed superior results to the visual analysis of the radiologist. When the radiologist associated both examinations (B-mode ultrasound and elastography), his visual analysis provided 87.10%, 55.77% and 0.714 of sensitivity, specificity and AUC, respectively. When we considered the result provided by the association between B-mode ultrasound and elastography images, the CADx system provided a comparative increase of about 7% of sensitivity and 17.2% of specificity, using the contour delimited by the most experienced radiologist. In addition, a positive influence was observed in the use of the computational tool by radiologists, since, on average, their sensitivity and specificity indexes also increased in relation to the conventional analysis, from 87.1% and 55.8% to 90.3% and 73.1%, respectively.

隨著智慧型手機的普及，在移動裝置上直接處理圖像的需求也大幅增加，故對於影像特徵描述子的要求，除了要表現出區域特徵的穩健性，同時也要維持良好的特徵比對效率與合理的儲存空間。過去所提出的區域影像特徵描述子建構方法之中，LIOP方法具有相當不錯的表現力，但其特徵描述子維度會隨著點群取樣數量的提高而以倍數增加，因此本研究提出Dynamic Local Intensity Order Relations (DLIOR)特徵描述子建構方法，利用LIOR方法探討點群中點與點之間的關係，減緩其維度增長幅度；透過動態設定像素差距門檻值，處理影像間像素差距分佈不均的問題，並使用線性轉換、點對歐幾里德距離等方式，重新定義描述子欄位的權重設定。經過實驗證實，DLIOR方法能夠使用比LIOP方法更少的維度空間，描述更多點群數的特徵資訊，並且具有更高的特徵比對能力。 / With the popularity of smart phones, the amounts of images being captured and processed on mobile devices have grown significantly in recent years. Image feature descriptors, which play crucial roles in recognition tasks, are expected to exhibit robust matching performance while at the same time maintain reasonable storage requirement. Among the local feature descriptors that have been proposed previously, local intensity order patterns (LIOP) demonstrated superior performance in many benchmark studies. As LIOP encodes the ranking relation in a point set (with N elements), however, its feature dimension increases drastically (N!) with the number of the neighboring sampling points around a pixel. To alleviate the dimensionality issue, this thesis presents a local feature descriptor by considering pairwise intensity relation in a pixel group, thereby reducing feature dimension to the order of C^N_2. In the proposed method, the threshold for assigning order relation is set dynamically according to local intensity distribution. Different weighting schemes, including linear transformation and Euclidean distance, have also been investigated to adjust the contribution of each pairing relation. Ultimately, the dynamic local intensity order relations (DLIOR) is devised to effectively encode intensity order relation of each pixel group. Experimental results indicate that DLIOR consumes less storage space than LIOP but achieves better feature matching performance using benchmark dataset.

This master’s thesis deals with an important part of computer vision field. Main focus of this thesis is on feature detectors and descriptors in an image. Throughout the thesis the simplest feature detectors like Moravec detector will be presented, building up to more complex detectors like MSER or FAST. The purpose of feature descriptors is in a mathematical description of these points. We begin with the oldest ones like SIFT and move on to newest and best performing descriptors like FREAK or ORB. The major objective of the thesis is comparison of presented methods on licence plate localization task.