Search

Hand segmentation is an integral part of many computer vision applications, especially gesture recognition. Training a classifier to classify pixels into hand or background using skin color as a feature is one of the most popular methods for this purpose. This approach has been highly restricted to simple hand segmentation scenarios since color feature alone provides very limited information for classification. Meanwhile there have been a rise of segmentation methods utilizing deep learning networks to exploit multi-layers of complex features learned from image data. Yet a deep neural network requires a large database for training and a powerful computational machine for operations due to its complexity in computations. In this work, the development of comprehensive features and optimized uses of these features with a randomized decision forest (RDF) classifier for the task of hand segmentation in uncontrolled indoor environments is investigated. Newly designed image features and new implementations are provided with evaluations of their hand segmentation performances. In total, seven image features which extract pixel or neighborhood related properties from color images are proposed and evaluated individually as well as in combination. The behaviours of feature and RDF parameters are also evaluated and optimum parameters for the scenario under consideration are identified. Additionally, a new dataset containing hand images in uncontrolled indoor scenarios was created during this work. It was observed from the research that a combination of features extracting color, texture, neighborhood histogram and neighborhood probability information outperforms existing methods for hand segmentation in restricted as well as unrestricted indoor environments using just a small training dataset. Computations required for the proposed features and the RDF classifier are light, hence the segmentation algorithm is suited for embedded devices equipped with limited power, memory, and

Trainable Superpixel Segmentation is a plug-in developed for the ImageJ platform that aims at providing its users with the ability to train models to segment images by classifying superpixels using region-based image features. This project provides an underlying library that can be used independently, a graphic interface for ease of use and an evaluation protocol of the efficacy of the library. The evaluation of the developed library was conducted through a ten-fold cross-validation and the results were compared with those of the Trainable Weka Segmentation library.

Trainable Superpixel Segmentation is a plug-in developed for the ImageJ platform that aims at providing its users with the ability to train models to segment images by classifying superpixels using region-based image features. This project provides an underlying library that can be used independently, a graphic interface for ease of use and an evaluation protocol of the efficacy of the library. The evaluation of the developed library was conducted through a ten-fold cross-validation and the results were compared with those of the Trainable Weka Segmentation library.

Trainable Superpixel Segmentation is a plug-in developed for the ImageJ platform that aims at providing its users with the ability to train models to segment images by classifying superpixels using region-based image features. This project provides an underlying library that can be used independently, a graphic interface for ease of use and an evaluation protocol of the efficacy of the library. The evaluation of the developed library was conducted through a ten-fold cross-validation and the results were compared with those of the Trainable Weka Segmentation library.

Staff-line removal is an important preprocessing stage for most optical music recognition systems. Common procedures to solve this task involve image processing techniques. In contrast to these traditional methods based on hand-engineered transformations, the problem can also be approached as a classification task in which each pixel is labeled as either staff or symbol, so that only those that belong to symbols are kept in the image. In order to perform this classification, we propose the use of convolutional neural networks, which have demonstrated an outstanding performance in image retrieval tasks. The initial features of each pixel consist of a square patch from the input image centered at that pixel. The proposed network is trained by using a dataset which contains pairs of scores with and without the staff lines. Our results in both binary and grayscale images show that the proposed technique is very accurate, outperforming both other classifiers and the state-of-the-art strategies considered. In addition, several advantages of the presented methodology with respect to traditional procedures proposed so far are discussed.

Staff-line removal is an important preprocessing stage for most optical music recognition systems. Common procedures to solve this task involve image processing techniques. In contrast to these traditional methods based on hand-engineered transformations, the problem can also be approached as a classification task in which each pixel is labeled as either staff or symbol, so that only those that belong to symbols are kept in the image. In order to perform this classification, we propose the use of convolutional neural networks, which have demonstrated an outstanding performance in image retrieval tasks. The initial features of each pixel consist of a square patch from the input image centered at that pixel. The proposed network is trained by using a dataset which contains pairs of scores with and without the staff lines. Our results in both binary and grayscale images show that the proposed technique is very accurate, outperforming both other classifiers and the state-of-the-art strategies considered. In addition, several advantages of the presented methodology with respect to traditional procedures proposed so far are discussed.

This work presents a novel approach to tackle the music staff removal. This task is devoted to removing the staff lines from an image of a music score while maintaining the symbol information. It represents a key step in the performance of most optical music recognition systems. In the literature, staff removal is usually solved by means of image processing procedures based on the intrinsics of music scores. However, we propose to model the problem as a supervised learning classification task. Surprisingly, although there is a strong background and a vast amount of research concerning machine learning, the classification approach has remained unexplored for this purpose. In this context, each foreground pixel is labelled as either staff or symbol. We use pairs of scores with and without staff lines to train classification algorithms. We test our proposal with several well-known classification techniques. Moreover, in our experiments no attempt of tuning the classification algorithms has been made, but the parameters were set to the default setting provided by the classification software libraries. The aim of this choice is to show that, even with this straightforward procedure, results are competitive with state-of-the-art algorithms. In addition, we also discuss several advantages of this approach for which conventional methods are not applicable such as its high adaptability to any type of music score.

This work presents a novel approach to tackle the music staff removal. This task is devoted to removing the staff lines from an image of a music score while maintaining the symbol information. It represents a key step in the performance of most optical music recognition systems. In the literature, staff removal is usually solved by means of image processing procedures based on the intrinsics of music scores. However, we propose to model the problem as a supervised learning classification task. Surprisingly, although there is a strong background and a vast amount of research concerning machine learning, the classification approach has remained unexplored for this purpose. In this context, each foreground pixel is labelled as either staff or symbol. We use pairs of scores with and without staff lines to train classification algorithms. We test our proposal with several well-known classification techniques. Moreover, in our experiments no attempt of tuning the classification algorithms has been made, but the parameters were set to the default setting provided by the classification software libraries. The aim of this choice is to show that, even with this straightforward procedure, results are competitive with state-of-the-art algorithms. In addition, we also discuss several advantages of this approach for which conventional methods are not applicable such as its high adaptability to any type of music score.

In this paper we empirically analyze the importance of sparsifying representations for classification purposes. We focus on those obtained by convolving images with linear filters, which can be either hand designed or learned, and perform extensive experiments on two important Computer Vision problems, image categorization and pixel classification. To this end, we adopt a simple modular architecture that encompasses many recently proposed models. The key outcome of our investigations is that enforcing sparsity constraints on features extracted in a convolutional architecture does not improve classification performance, whereas it does so when redundancy is artificially introduced. This is very relevant for practical purposes, since it implies that the expensive run-time optimization required to sparsify the representation is not always justified, and therefore that computational costs can be drastically reduced.

The task of segmenting the lung fields, the heart, and the clavicles in standard posterior-anterior chest radiographs is considered. Three supervised segmentation methods are compared: active shape models, active appearance models, both first proposed by Cootes et al. and a multi-resolution pixel classification method that employs a multi-scale filter bank of Gaussian derivatives and a k-nearest-neighbors classifier. The methods have been tested on a publicly available database of 247 chest radiographs, in which all objects have been manually segmented by two human observers. A parameter optimization for active shape models is presented and it is shown that this optimization improves performance significantly. It is demonstrated that the standard active appearance model scheme performs poorly, but large improvements can be obtained by including areas outside the objects into the model. For lung field segmentation, all methods perform well, with pixel classification giving the best results: a paired t-test showed no significant performance difference between pixel classification and an independent human observer. For heart segmentation, all methods perform comparably, but significantly worse than a human observer. Clavicle segmentation is a hard problem for all methods; best results are obtained with active shape models, but human performance is substantially better. As an application, the cardio-thoracic ratio is computed automatically from the segmentation results. Bland and Altman plots indicate that all methods perform well when compared to the gold standard, with confidence intervals from pixel classification and active appearance modeling very close to those of a human observer. All results, including the manual segmentations, have been made publicly available to facilitate future comparative studies.

The task of segmenting the lung fields, the heart, and the clavicles in standard posterior-anterior chest radiographs is considered. Three supervised segmentation methods are compared: active shape models, active appearance models, both first proposed by Cootes et al. and a multi-resolution pixel classification method that employs a multi-scale filter bank of Gaussian derivatives and a k-nearest-neighbors classifier. The methods have been tested on a publicly available database of 247 chest radiographs, in which all objects have been manually segmented by two human observers. A parameter optimization for active shape models is presented and it is shown that this optimization improves performance significantly. It is demonstrated that the standard active appearance model scheme performs poorly, but large improvements can be obtained by including areas outside the objects into the model. For lung field segmentation, all methods perform well, with pixel classification giving the best results: a paired t-test showed no significant performance difference between pixel classification and an independent human observer. For heart segmentation, all methods perform comparably, but significantly worse than a human observer. Clavicle segmentation is a hard problem for all methods; best results are obtained with active shape models, but human performance is substantially better. As an application, the cardio-thoracic ratio is computed automatically from the segmentation results. Bland and Altman plots indicate that all methods perform well when compared to the gold standard, with confidence intervals from pixel classification and active appearance modeling very close to those of a human observer. All results, including the manual segmentations, have been made publicly available to facilitate future comparative studies.

This work presents a study of three important issues of the color pixel classification approach to skin segmentation: color representation, color quantization, and classification algorithm. Our analysis of several representative color spaces using the Bayesian classifier with the histogram technique shows that skin segmentation based on color pixel classification is largely unaffected by the choice of the color space. However, segmentation performance degrades when only chrominance channels are used in classification. Furthermore, we find that color quantization can be as low as 64 bins per channel, although higher histogram sizes give better segmentation performance. The Bayesian classifier with the histogram technique and the multilayer perceptron classifier are found to perform better compared to other tested classifiers, including three piecewise linear classifiers, three unimodal Gaussian classifiers, and a Gaussian mixture classifier.

This work presents a study of three important issues of the color pixel classification approach to skin segmentation: color representation, color quantization, and classification algorithm. Our analysis of several representative color spaces using the Bayesian classifier with the histogram technique shows that skin segmentation based on color pixel classification is largely unaffected by the choice of the color space. However, segmentation performance degrades when only chrominance channels are used in classification. Furthermore, we find that color quantization can be as low as 64 bins per channel, although higher histogram sizes give better segmentation performance. The Bayesian classifier with the histogram technique and the multilayer perceptron classifier are found to perform better compared to other tested classifiers, including three piecewise linear classifiers, three unimodal Gaussian classifiers, and a Gaussian mixture classifier.

Many tasks in computer vision lead to combinatorial optimization problems. Automatic image partitioning is one of the most important examples in this context: whether based on some prior knowledge or completely unsupervised, we wish to find coherent parts of the image. However, the inherent combinatorial complexity of such problems often prevents to find the global optimum in polynomial time. For this reason, various approaches have been proposed to find good approximative solutions for image partitioning problems. As an important example, we will first consider different spectral relaxation techniques: based on straightforward eigenvector calculations, these methods compute suboptimal solutions in short time. However, the main contribution of this thesis is to introduce a novel optimization technique for discrete image partitioning problems which is based on a semidefinite programming relaxation. In contrast to approximation methods employing annealing algorithms, this approach involves solving a convex optimization problem, which does not suffer from possible local minima. Using interior point techniques, the solution of the relaxation can be found in polynomial time, and without elaborate parameter tuning. High quality solutions to the original combinatorial problem are then obtained with a randomized rounding technique. The only potential drawback of the semidefinite relaxation approach is that the number of variables of the optimization problem is squared. Nevertheless, it can still be applied to problems with up to a few thousand variables, as is demonstrated for various computer vision tasks including unsupervised segmentation, perceptual grouping and image restoration. Concerning problems of higher dimensionality, we study two different approaches to effectively reduce the number of variables. The first one is based on probabilistic sampling: by considering only a small random fraction of the pixels in the image, our semidefinite relaxation method can be app, Viele Bildverarbeitungsaufgaben lassen sich auf kombinatorische Optimierungsprobleme zurückführen. Eines der wichtigsten Beispiele in diesem Kontext ist die automatische Zerlegung von Bildern in kohärente Bestandteile, sei es unter Zuhilfenahme von Vorwissen oder völlig unüberwacht. Allerdings erlaubt es die hohe Komplexität derartiger Probleme häufig nicht, optimale Lösungen in polynomieller Zeit zu berechnen. Aus diesem Grund wurden verschiedenartige Verfahren entwickelt, um gute Näherungslösungen für Bildpartitionierungsprobleme zu bestimmen. Als wichtiges Beispiel vergleichen wir zunächst mehrere spektrale Relaxations-Methoden, welche solche Approximationen mit Hilfe von speziellen Eigenvektor-Berechnungen ermitteln. Der wesentliche Beitrag dieser Arbeit besteht jedoch darin, ein neuartiges Optimierungsverfahren zur diskreten Bildpartitionierung vorzustellen. Wir verwenden dazu einen Relaxationsansatz, der letztlich ein spezielles konvexes Optimierungsproblem liefert, welches mittels semidefiniter Programmierung gelöst werden kann. Im Gegensatz zu anderen Näherungsverfahren, die beispielsweise auf Annealing-Algorithmen beruhen, besteht somit keine Gefahr, in einem lokalen Minimum zu landen. Außerdem kann die Lösung eines semidefiniten Programms ohne aufwändige Parameteroptimierung mit Interior-Point-Methoden in polynomieller Zeit bestimmt werden. Qualitativ hochwerige diskrete Lösungen des Originalproblems lassen sich anschließend mit Hilfe einer probabilistischen Rundungstechnik ermitteln. Der einzige potenzielle Nachteil der semidefiniten Relaxation besteht darin, dass eine Quadrierung der Variablenanzahl notwendig ist. Nichtsdestotrotz lassen sich Probleme mit bis zu einigen tausend Variablen zufriedenstellend bearbeiten, wie wir anhand unterschiedlicher Bildverarbeitungsaufgaben aus der unüberwachten Segmentierung, perzeptuellen Gruppierung oder Bildrekonstruktion demonstrieren werden. Für Problemstellungen höherer Dimension untersuchen wir zwei verschiedene

An unsupervised color image segmentation algorithm is presented, using a Markov random field (MRF) pixel classification model. We propose a new method to estimate initial mean vectors effectively even if the histogram does not have clearly distinguishable peaks. The only parameter supplied by the user is the number of classes. (C) 2001 Elsevier Science B.V. All rights reserved.

An unsupervised color image segmentation algorithm is presented, using a Markov random field (MRF) pixel classification model. We propose a new method to estimate initial mean vectors effectively even if the histogram does not have clearly distinguishable peaks. The only parameter supplied by the user is the number of classes. (C) 2001 Elsevier Science B.V. All rights reserved.

An unsupervised color image segmentation algorithm is presented, using a Markov random field (MRF) pixel classification model. We propose a new method to estimate initial mean vectors effectively even if the histogram does not have clearly distinguishable peaks. The only parameter supplied by the user is the number of classes. (C) 2001 Elsevier Science B.V. All rights reserved.

An inadequate concept of how corresponding points relate to one another on dissimilar images has a greater effect than exposure geometry or data collection on registration problems in stereo photogrammetry. Conventional correlation, or one of its relatives, is the measure of similarity used in all automated stereo correlation systems. Correlation, a measure of the linear dependence between two sets of data, is an inadequate measure when there is less than, or more than, a moderate amount of image structure at and around points selected for image matching. The existence of structure should be recognized and utilized in an appropriate manner for image matching. Similarly, the absence of structure should be recognized, and the surrounding imagery should be used to complete matches where it is possible. The concurrent determination of what a pixel is, as well as where it is, can alleviate much of the registration problem. A variety of features including point-density data, texture, and edges, as well as existing cartographic knowledge, can be combined and organized through rules in order to more completely describe a point. The overall throughput of the compilation process will be improved in both time and accuracy if those functions which tend to support one another are concurrently, rather than sequentially, performed. If the compilation process takes place in image space, then the image matching operation as well as the other feature extraction operations, can be ordered by the data processing manager to best suit the function of the process. (Author)