Refine
Document Type
- Doctoral Thesis (2)
Language
- English (2)
Has Fulltext
- yes (2)
Is part of the Bibliography
- no (2)
Keywords
- Machine learning (2) (remove)
Institute
The segmentation of volumetric datasets, i.e., the partitioning of the data into disjoint sub-volumes with the goal to extract information about these regions,is a difficult problem and has been discussed in medical imaging for decades.
Due to the ever-increasing imaging capabilities, in particular in X-ray computed tomography (CT) or magnetic resonance imaging, segmentation in industrial applications also gains interest.
Especially in industrial applications the generated datasets increase in size.
Hence, most applications apply well-known techniques in a 2+1-dimensional manner,i.e., they apply image segmentation procedures on each slice separately and track the progress along the axis of the volume in which the slices are stacked on.
This discards the information on preceding or subsequent slices, which is often assumed to be nearly identical. However, in the industrial context this might prove wrong since industrial parts might change their appearance significantly over the course of even a few slices.
Moreover, artifacts can further distort the content of the slices.
Therefore, three-dimensional processing of voxel volumes has to be preferred, which induces constraints upon the segmentation procedures. For example, they must not consider global information as it is usually not feasible in big scans to compute them efficiently.
Yet another frequent problem is that applications focus on individual parts only and algorithms are tailored to that case. Most prominent medical segmentation procedures do so by applying methods to specifically find the liver and only the liver of a patient, for example.
The implication is that the same method then cannot be applied to find other parts of the scan and such methods have to be designed individually for any object to be segmented.
Flexible segmentation methods are needed too specifically when partitioning unique scans. We define a unique scan to be a voxel dataset for which no comparable volume exists.
Classical examples include the use case of cultural heritage where not only the objects themselves are unique but also scan parameters are optimized to obtain the best image quality possible for that specific scan.
This thesis aims at introducing novel methods for voxelwise classifications based on local geometric features.
The latter are computed from local environments around each voxel and extract information in similar ways as humans do, namely by observing their similarity to geometric or textural primitives.
These features serve as the foundation to learning the proposed voxelwise classifiers and to discriminate between segmented and unsegmented voxels.
On the one hand, they perform fully automated clustering of volumes for which a representative random sample is extracted first.
On the other hand, a set of segmenting classifiers can be trained from few seed voxels, i.e., volume elements for which a domain expert marked if they belong to the components that shall be segmented. The interactive selection offers the advantage that no completely labeled voxel volumes are necessary and hence that unique scans of objects can be segmented for which no comparable scans exist.
Overall, it will be shown that all proposed segmentation methods are effectively of linear runtime with respect to the number of voxels in the volume. Thus, voxel volumes without size restrictions can be segmented in an efficient linear pass through the volume.
Finally, the segmentation performance is evaluated on selected datasets which shows that the introduced methods can achieve good results on scans from a broad variety of domains for both small and big voxel volumes.
In the last decade, crowdsourcing has proved its ability to address large scale data collection tasks, such as labeling large data sets, at a low cost and in a short time. However, the performance and behavior variability between workers as well as the variability in task designs and contents, induce an unevenness in the quality of the produced contributions and, thus, in the final output quality. In order to maintain the effectiveness of crowdsourcing, it is crucial to control the quality of the contributions. Furthermore, maintaining the efficiency of crowdsourcing requires the time and cost overhead related to the quality control to be at its lowest. While effective, current quality control techniques such as contribution aggregation, worker selection, context-specific reputation systems, and multi-step workflows, suffer from fairly high time and budget overheads and from their dependency on prior knowledge about individual workers.
In this thesis, we address this challenge by leveraging the similarity between completed and incoming tasks as well as the correlation between the worker declarative profiles and their performance in previous tasks in order to perform an efficient task-aware worker selection. To this end, we propose CAWS (Context AwareWorker Selection) method which operates in two phases; in an offline phase, completed tasks are clustered into homogeneous groups for each of which the correlation with the workers declarative profile is learned. Then, in the online phase, incoming tasks are matched to one of the existing clusters and the correspondent, previously inferred profile model is used to select the most reliable online workers for the given task. Using declarative profiles helps eliminate any probing process, which reduces the time and the budget while maintaining the crowdsourcing quality. Furthermore, the set of completed tasks, when compared to a probing task split, provides a larger corpus from which a more precise profile model can be learned. This translates to a better selection quality, especially for harder tasks.
In order to evaluate CAWS, we introduce CrowdED (Crowdsourcing Evaluation Dataset), a rich dataset to evaluate quality control methods and quality-driven task vectorization and clustering. The generation of CrowdED relies on a constrained sampling approach that allows to produce a task corpus which respects both, the budget and type constraints. Beside helping in evaluating CAWS, and through its generality and richness, CrowdED helps in plugging the benchmarking gap present in the crowdsourcing quality control community.
Using CrowdED, we evaluate the performance of CAWS in terms of the quality of the worker selection and in terms of the achieved time and budget reduction. Results shows the following: first, automatic grouping is able to achieve a learning quality similar to job-based grouping. And second, CAWS is able to outperform the state-of-the-art profile-based worker selection when it comes to quality. This is especially true when strong budget and time constraints are present on the requester side.
Finally, we complement our work by a software contribution consisting of an open source framework called CREX (CReate Enrich eXtend). CREX allows the creation, the extension and the enrichment of crowdsourcing datasets. It provides the tools to vectorize, cluster and sample a task corpus to produce constrained task sets and to automatically generate custom crowdsourcing campaign sites.