Category Independent Object Proposals Using Quantum Superposition
Malik, Junaid (2017)
2017
Master's Degree Programme in Information Technology
Tieto- ja sähkötekniikan tiedekunta - Faculty of Computing and Electrical Engineering
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2017-03-08
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201701191061
https://urn.fi/URN:NBN:fi:tty-201701191061
Tiivistelmä
A vast amount of digital images and videos are continually being generated and shared across the Internet. An important step towards utilizing this ‘big data’ and deducing meaningful information from its visual contents, is to detect the presence of objects belonging to a particular class in digital images. Earlier computer vision algorithms devised for this purpose exhaustively search the entire image space for detecting objects belonging to a particular class. Object proposals aim to reduce this search space by proposing probable locations of objects in the image beforehand. This paves the way for efficiently using more computationally expensive and sophisticated detection algorithms.
Conventional approaches to generating object proposals have revolved around learning a scoring function from the characteristics of objects in ground truth annotations of images. In this thesis, we propose a novel category independent proposal generation framework that is unsupervised and inspired by the psycho-visual analysis of human visual system where the search for objects gradually transitions from the most salient parts of a scene to comparatively non-salient regions. We use a state-of-the-art visual saliency estimation technique which proposes a unique relationship between spectral clustering and quantum mechanics. We improve this method by exploiting for the first time, the quantum superposition principle, to extend the search of objects beyond the salient ones. We also propose an unsupervised scoring strategy that does not incorporate any prior information about the spatial, color or textural features of objects.
Experimental results have proved that our proposed methodology achieves comparable results with the contemporary state-of-the-art methods. Our unsupervised scoring strategy is shown to outperform, in some cases, the supervised frameworks employed by other methods. Moreover, it also enables us to achieve a three-fold decrease in the number of proposals while keeping the loss of recall to less than 3%. The success of our proposed methodology opens the door to a research direction where quantum mechanical principles can be utilized to enable computer vision algorithms to find objects in digital images without having any prior knowledge about them.
Conventional approaches to generating object proposals have revolved around learning a scoring function from the characteristics of objects in ground truth annotations of images. In this thesis, we propose a novel category independent proposal generation framework that is unsupervised and inspired by the psycho-visual analysis of human visual system where the search for objects gradually transitions from the most salient parts of a scene to comparatively non-salient regions. We use a state-of-the-art visual saliency estimation technique which proposes a unique relationship between spectral clustering and quantum mechanics. We improve this method by exploiting for the first time, the quantum superposition principle, to extend the search of objects beyond the salient ones. We also propose an unsupervised scoring strategy that does not incorporate any prior information about the spatial, color or textural features of objects.
Experimental results have proved that our proposed methodology achieves comparable results with the contemporary state-of-the-art methods. Our unsupervised scoring strategy is shown to outperform, in some cases, the supervised frameworks employed by other methods. Moreover, it also enables us to achieve a three-fold decrease in the number of proposals while keeping the loss of recall to less than 3%. The success of our proposed methodology opens the door to a research direction where quantum mechanical principles can be utilized to enable computer vision algorithms to find objects in digital images without having any prior knowledge about them.