dc.contributor.author |
Elia, Gaby Ebrahim. |
dc.date.accessioned |
2013-10-02T09:22:04Z |
dc.date.available |
2013-10-02T09:22:04Z |
dc.date.issued |
2012 |
dc.identifier.uri |
http://hdl.handle.net/10938/9499 |
dc.description |
Thesis (M.E.)--American University of Beirut, Department of Electrical and Computer Engineeering, 2012. |
dc.description |
Advisor : Dr. Mohammad Mansour, Associate Professor, Electrical and Computer Engineering--Committee Members : Dr. Mohammad Adnan Alaoui, Professor, Electrical and Computer Engineering ; Dr. Hazem Hajj, Associate Professor, Electrical and Computer Engineering. |
dc.description |
Includes bibliographical references (leaves 191-194) |
dc.description.abstract |
Among the many object recognition models in computer vision, the recently proposed biologically inspired models, based on models of the visual cortex, have assumed significance developments.These models, called HMAX (Hierarchical Model and X), apply Gabor filters at different positions and scales and perform alternate template matching(S layers) and max pooling operations (C layers) to build feature complexity and position-scale invariance. Gabor filters as models of V1 neurons have received considerable attention in image processing for applications such as texture analysis, iris, face recognition, object detection, and finger print recognition. While Gabor filter functions extract simple orientation features, HMAX and models based on it derive complex structural features built up on simple Gabor filters. These models involve tremendous amounts of computation and have not been incorporated into real-world applications which need to operate at real-time.Based on that we are interested in enhancing this model by applying cloning method in order to improve the confidence accuracy for classification and detection, then accelerating the computations in the HMAX layers by building a reconfigurable hardware accelerator for the time-consuming S-C stages of the HMAX model and evaluate its performance on an FPGA. The S2 stage involves variable sized template matching on a large number of templates or patches on a multi-scale feature representation of an image. Our final model is tested on the Caltech 101 object categories (which contain 9145 images), and UIUC car data (1050 images) in order to train the model. |
dc.format.extent |
xv, 194 leaves : ill. ; 30 cm. |
dc.language.iso |
eng |
dc.relation.ispartof |
Theses, Dissertations, and Projects |
dc.subject.classification |
ET:005747 AUBNO |
dc.subject.lcsh |
Computer vision. |
dc.subject.lcsh |
Computer algorithms. |
dc.subject.lcsh |
Computer architecture. |
dc.subject.lcsh |
Image processing. |
dc.subject.lcsh |
Kernel functions. |
dc.subject.lcsh |
Software engineering. |
dc.title |
High speed architectures for accelerating neuromorphic vision algorithms. |
dc.type |
Thesis |
dc.contributor.department |
American University of Beirut. Faculty of Engineering and Architecture. Department of Electrical and Computer Engineering. |