Utilize este identificador para referenciar este registo: http://hdl.handle.net/10400.1/3639
Título: Object categorization using biological models
Autor: Vieira, Ana Milene Mestre
Orientador: Rodrigues, J. M. F.
Data de Defesa: 2013
Resumo: Humans are naturals at categorizing objects, i.e., at dividing them into groups depending on their features and surroundings. We do it easily and in real-time. Additionally, our Human Visual System (HVS) is the only one reliable for object detection, categorization and recognition; the latter events take place in the visual cortex, being object recognition achieved around 150-200ms, and occurring also a categorization-specific activation in prefrontal cortex before or around 100ms. This provides one of the evidences which substantiate that categorization is a more bottom-up process than recognition. Visual cortical area V1 is composed - among others - by simple and complex cells which are adjusted to different spatial frequencies (scales), orientations and disparity. These cell‟s responses were used to build a model for events detection in V1; these events are classified by type - lines and edges – and polarity - positive and negative. Being the goal of this thesis to develop a cortical model for object categorization - inspired in the HVS and based on 2D object views -, the V1 multi-scale events generated by the former model were used to accomplish that goal. In the developed categorization model the final category attributed to an object is the convergence of three similarity concepts which define in different ways the resemblance degree between an object and a certain category; the resemblance degree is therefore accomplished by comparing the V1 events between templates and objects. The resemblance degree or similarity percentage was calculated (a) on the first concept as the quotient between the number of common events between object and category templates (considering type and polarity) in all scales, and the number of object‟s events in all scales; (b) on the second concept the similarity percentage was calculated as the quotient between the number of common events between object and category templates (not considering type nor polarity) in all scales, and the number of object‟s events in all scales; (c) finally, on the third concept this ratio was calculated as the quotient between the number of common events between object and category templates (considering type and polarity) in all scales, and the category‟s “events number” in all scales. The final category assigned to an object is then (1st) a category on which the three concepts agree on and (2nd) the best scored one. For the proof of concept a database composed by 8 different categories and 10 objects per category was used; left and right profile views were chosen to represent each object. Regarding the 80 results obtained by categorizing 40 objects in both views, an average categorization success rate of 93.75% was accomplished, being 92.50% the success rate achieved for left profile, and 95.00% the one achieved for right profile; even each of the miscategorized images was attributed a category which is similar to its true one. In order to conclude the proof of concept, the model was also tested in terms of small invariance to rotation, scale and noise, having been then achieved high categorization success rates (above 82%).
Descrição: Dissertação de mest., Engenharia Elétrica e Eletrónica (Tecnologias da Informação e Telecomunicação), Instituto Superior de Engenharia, Univ. do Algarve, 2013
Peer review: yes
URI: http://hdl.handle.net/10400.1/3639
Designação: Mestrado em Engenharia Elétrica e Eletrónica. Especialização em Tecnologias da Informação e telecominucação
Aparece nas colecções:UA01-Teses

Ficheiros deste registo:
Ficheiro Descrição TamanhoFormato 
Tese_final.pdf3,63 MBAdobe PDFVer/Abrir


FacebookTwitterDeliciousLinkedInDiggGoogle BookmarksMySpace
Formato BibTex MendeleyEndnote Degois 

Todos os registos no repositório estão protegidos por leis de copyright, com todos os direitos reservados.