Full text
| |
| Author(s): |
Paulo Eduardo Rauber
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2017-02-20 |
| Examining board members: |
Alexandre Xavier Falcão;
Henk Wolter Broer;
Jirí Kosinka;
Michael Biehl;
Gerard Rudolf Renardel de Lavalette;
Jarke Jan van Wijk;
Ricardo da Silva Torres
|
| Advisor: | Johannes Bernardus Theodorus Maria Roerdink; Pedro Jussieu de Rezende; Alexandre Xavier Falcão |
| Abstract | |
We define image analysis as the field of study concerned with extracting information from images. This field is immensely important for commercial and interdisciplinary applications. The overarching goal behind the work presented in this thesis is enabling user interaction during several tasks related to image analysis: image segmentation, feature selection, and image classification. In this context, enabling user interaction refers to providing mechanisms that allow humans to assist machines in tasks that are difficult to automate. Such tasks are very common in image analysis. Concerning image segmentation, we propose a new interactive technique that combines superpixels with the image foresting transform. The main advantage of our proposed technique is enabling faster interactive segmentation of large images, although it also enables potentially richer feature extraction. Our experiments show that our technique is at least as effective as its pixel-based counterpart. In the context of feature selection and image classification, we propose a new interactive visualization system that combines feature space exploration (based on dimensionality reduction) with automatic feature scoring. This visualization system aims to provide insights that lead to the development of effective feature sets for image classification. The same system can also be applied to select features for image segmentation and (general) pattern classification, although these tasks are not our focus. We present use cases that show how this system may provide a kind of qualitative feedback about image classification systems that would be very difficult to obtain by other (non-visual) means. We also show how our proposed interactive visualization system can be adapted to explore intermediary computational results of artificial neural networks. Such networks currently achieve state-of-the-art results in many image classification applications. Through use cases involving traditional benchmark datasets, we show that our system may enable insights about how a network operates that lead to improvements along the classification pipeline. Because the parameters of an artificial neural network are typically adapted iteratively, visualizing its intermediary computational results can be seen as a time-dependent task. Motivated by this, we propose a new time-dependent dimensionality reduction technique that enables the reduction of apparently unnecessary changes in results due to small changes in the data (temporal coherence). Preliminary experiments show that this technique is effective in enforcing temporal coherence (AU) | |
| FAPESP's process: | 12/24121-9 - Multi-label annotation of natural image databases with minimal supervision |
| Grantee: | Paulo Eduardo Rauber |
| Support Opportunities: | Scholarships in Brazil - Doctorate (Direct) |