Josef Pauli - Academia.edu (original) (raw)

Papers by Josef Pauli

Signal Processing, Pattern Recognition and Applications, 2010

Robust time-to-contact calculation belongs to the most desirable techniques in the field of auton... more Robust time-to-contact calculation belongs to the most desirable techniques in the field of autonomous robot navigation. Using only image measurements it provides a method to determine when contact with a visible object will be made. However the computation of the time-to- contact values is very sensitive to noisy measurements of feature positions in a image. Instead of developing a new feature extraction and tracking algorithm this paper presents an approach which deals with the inaccurate measurements. It is based on the here derived equations which describe the process how a feature diverges from the focus of expansion. The results presented testify the stability and the robustness of this approach.

With the objective of building robots that accompany humans in daily life, it might be favourable... more With the objective of building robots that accompany humans in daily life, it might be favourable that such robots act humanlike so that humans are able to predict their behaviour without effort. Decision making is one crucial aspect of daily life. As Damasio demonstrated, human decisions are often based on emotions. Earlier work thus developed a decision making framework for artificial intelligent systems based on Damasio's Somatic Marker Hypothesis and revealed that overall, the decisions made by an artificial agent resemble those of human players. This paper enhances this work in so far that a detailed evaluation of the first 30 decisions made by the modelled agent during this gambling task was done by human subjects. Therefore 26 human participants were recruited who had to evaluate different graphical outputs that visualized the course of the Iowa Gambling Task played by either a modelled agent or a human. The results revealed that participants tend to categorize the cours...

Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies, 2021

Proceedings of the 9th International Conference on Computer Vision Theory and Applications, 2014

In this paper we study the optimization process of an object classification task for an image-bas... more In this paper we study the optimization process of an object classification task for an image-based steel quality measurement system. The goal is to distinguish hollow from solid defects inside of steel samples by using texture and shape features of reconstructed 3D objects. In order to optimize the classification results we propose a holistic machine learning framework that should automatically answer the question "How well do state-of-the-art machine learning methods work for my classification problem?" The framework consists of three layers, namely feature subset selection, feature transform and classifier which subsequently reduce the data dimensionality. A system configuration is defined by feature subset, feature transform function, classifier concept and corresponding parameters. In order to find the configuration with the highest classifier accuracies, the user only needs to provide a set of feature vectors and ground truth labels. The framework performs a totally data-driven optimization using partly heuristic grid search. We incorporate several popular machine learning concepts, such as Principal Component Analysis (PCA), Support Vector Machines (SVM) with different kernels, random trees and neural networks. We show that with our framework even non-experts can automatically generate a ready for use classifier system with a significantly higher accuracy compared to a manually arranged system.

This paper presents a method to support point cloud registration with color information. For this... more This paper presents a method to support point cloud registration with color information. For this purpose we integrate L?a?b? color space information into the Generalized Iterative Closest Point (GICP) algorithm, a state-of-the-art Plane-To-Plane ICP variant. A six-dimensional k-d tree based nearest neighbor search is used to match corresponding points between the clouds. We demonstrate that the additional effort in general does not have an immoderate impact on the runtime, since the number of iterations can be reduced. The influence on the estimated 6 DoF transformations is quantitatively evaluated on six different datasets. It will be shown that the modified algorithm can improve the results without needing any special parameter adjustment.

In single-sensor digital imaging a color filter array, that is overlaid onto the image sensor, ma... more In single-sensor digital imaging a color filter array, that is overlaid onto the image sensor, makes color images possible. Incident light rays become band-limited and each sensor element captures either red, green or blue light. Interpolating the missing two color components for each pixel location is known as demosaicing. This paper proposes to firstly derive an estimated luminance image by low-pass filtering the original mosaiced sensor image. In a second step a deconvolution technique re-sharpenes the blurred luminance approximation, so that it has the same spatial resolution as the original - but bandpassed - sensor image. Using the high-resolution luminance approximation the partial RGB colors from the mosaiced sensor image are transformed into a different color space that is more suitable for color interpolation. The new color space consists of least correlated color data, so that intra-channel interpolation errors have a reduced impact on inter-channel alignment, and therefo...

Various segmentation approaches in medical image processing, such as Level Sets, Active Shape Mod... more Various segmentation approaches in medical image processing, such as Level Sets, Active Shape Models, and Active Appearance Models require initial localization of the structure of interest (SOI). In this work we present a novel fully automatic model initialization approach in MRI, that is applicable for structures that are mostly convex in the axial plane. We propose a training model, namely the Polar Appearance Model, that encapsulates both the transition from the structure of interest to its vicinity in polar space and the intensity distribution within the structure in euclidean space. We present our approach on the example of femoral model initialization in MRI and compare our results to a standard voxel-based registration approach that allows similarity transformations.

– This thesis approaches the problem of detecting osteoporosis on X-ray images of different pati... more – This thesis approaches the problem of detecting osteoporosis on X-ray images of different patients by means of machine learning. Osteoporosis has become one of the western world’s most imminent conditions [1] for elderly people, both male and female, as it causes a reduction in bone strength thus leading to a higher risk of broken bones [2]. While effective treatment is still largely unknown, early detection of osteoporosis can help prevent or at least delay the disease by means of a proper diet, adequate exercises, medications and also lifestyle changes [3]. In many cases osteoporosis first comes to the patient’s notice when a minor injury, like a fall, causes a fracture. This however indicates an advanced stage of the disease, in which prevention or retardation efforts are ineffective. The necessity for early detection is obvious, yet current methods for detection are solely applicable if there is reasonable suspicion that the patient in fact suffers from the disease. The sta...

The problem of tone mapping for HDR (high dynamic range) to LDR (low dynamic range) conversion is... more The problem of tone mapping for HDR (high dynamic range) to LDR (low dynamic range) conversion is introduced by a unified framework considering all the usual processing steps. Then the specific problem of single-shot HDR is outlined where special emphasis is taken on the effect of the greater noise floor of those images when compared to the usual exposure bracketing approach to HDR. We herein tailor the popular tone mapping operators proposed by Reinhard for single-shot HDR. A region-based approach for preprocessing any HDR image in order to increase SNR and perceptual sharpness is introduced as an extension to our initial tone mapping framework. The results are compared with respect to specially developed baseline tone mappers and an extensive subjective evaluation is performed.

In this paper, we present a 2D deep multitask learning approach for the segmentation of small str... more In this paper, we present a 2D deep multitask learning approach for the segmentation of small structures on the example of avascular necrosis of the femoral head (AVNFH) in MRI. It consists of one joint encoder and three separate decoder branches, each assigned to its own objective. We propose using a reconstruction task to initially pre-train the encoder and shift the objective towards a second necrosis segmentation task in a reconstruction-dependent loss adaptation manner. The third branch deals with the rough localization of the topographical neighborhood of possible femoral necrosis areas. Its output is used to emphasize the roughly approximated location of the segmentation branch’s output. The evaluation of the segmentation performance of our architecture on coronal T1-weighted MRI volumes shows promising improvements compared to a standard U-Net implementation.

Medical Image Analysis, 2021

Segmentation of abdominal organs has been a comprehensive, yet unresolved, research field for man... more Segmentation of abdominal organs has been a comprehensive, yet unresolved, research field for many years. In the last decade, intensive developments in deep learning (DL) have introduced new state-of-the-art segmentation systems. Despite outperforming the overall accuracy of existing systems, the effects of DL model properties and parameters on the performance is hard to interpret. This makes comparative analysis a necessary tool to achieve explainable studies and systems. Moreover, the performance of DL for emerging learning approaches such as cross-modality and multi-modal tasks have been rarely discussed. In order to expand the knowledge in these topics, CHAOS-Combined (CT-MR) Healthy Abdominal Organ Segmentation challenge has been organized in the IEEE International Symposium on Biomedical Imaging (ISBI), 2019, in Venice, Italy. Despite a large number of the previous abdomen related challenges, the majority of which are focused on tumor/lesion detection and/or classification with a single modality, CHAOS provides This work is supported by Scientific and Technological Research Council of Turkey (TUBITAK) ARDEB-EEEAG under grant number 116E133.

Proceedings of the 3rd International Conference on Pattern Recognition Applications and Methods, 2014

Lecture Notes in Computer Science

Informatik aktuell, 2016

By this contribution we tackle the challenge of extracting fuzzy, curvilinear fine structures fro... more By this contribution we tackle the challenge of extracting fuzzy, curvilinear fine structures from medical image data. The dissection membrane represents a highly tortuous fine structure within aortas of dissection patients. Due to its variability in topology and morphology, extraction by any assumed shape priors is deemed to fail. Based on the response of 3D/2D phase congruency filter, we select a segment of high significance in order to remove false positives within each CTA slice. Multicriterial, greedy tracking serves for membrane completion, while a inter-slice grouping algorithm performs detection of global outliers. Erroneous slice results are replaced using sampled membrane segments from adjacent slices. Our proposed algorithm not only improves the membrane segmentation by up to 32% when compared to stand-alone usage of local phase features, but also enables separation of the true and false lumen.

Proceedings of the 10th International Conference on Computer Vision Theory and Applications, 2015

The development of image-based object recognition systems with the desired performance is-still-a... more The development of image-based object recognition systems with the desired performance is-still-a challenging task even for experts. The properties of the object feature representation have a great impact on the performance of any machine learning algorithm. Manifold learning algorithms like e.g. PCA, Isomap or Autoencoders have the potential to automatically learn lower dimensional and more useful features. However, the interplay of features, classifiers and hyperparameters is complex and needs to be carefully tuned for each learning task which is very time-consuming, if it is done manually. This paper uses a holistic optimization framework with feature selection, multiple manifold learning algorithms, multiple classifier concepts and hyperparameter optimization to automatically generate pipelines for image-based object classification. An evolutionary algorithm is used to efficiently find suitable pipeline configurations for each learning task. Experiments show the effectiveness of the proposed representation and classifier tuning on several high-dimensional object recognition datasets. The proposed system outperforms other state-of-the-art optimization frameworks.

Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 2018