Retrieval of human facial images based on visual content and semantic description / Ahmed Abdu Ali Alattab

Alattab, Ahmed Abdu Ali (2013) Retrieval of human facial images based on visual content and semantic description / Ahmed Abdu Ali Alattab. PhD thesis, University of Malaya.

[img]
Preview
PDF (Full Text)
Download (9Mb) | Preview

    Abstract

    The significant increase in the huge collections of digital images and videos that need to be managed has led to the requirement for efficient methods for the archival and retrieval of these images. Facial images have gained its importance amongst these digital images due to its use in various aspects of life such as, in airports, law enforcement applications, security systems and automated surveillance applications. The basic content-based image retrieval (CBIR) system used for the general task of image retrieval is not effective with facial images, especially when the query is in some form of user descriptions. The current CBIR is based on low-level features such as color, texture, shape, and eigenfaces thus it cannot capture the semantic aspects of a facial image. Humans by nature tend to use semantic descriptions (high-level feature) when describing what they are looking for, and they normally encounter difficulties when using descriptions based on low-level features. This is because human beings normally perceive facial images and compare their similarities using high-level features such as gender, race, age and the rating scale of the facial traits and thus cannot relate these high-level semantic concepts directly to low-level features. In this research, we propose a semantic content-based facial image retrieval technique (SCBFIR) that incorporates multiple visual features with semantic features to increase the accuracy of the facial image retrieval and to reduce the semantic gap between the high-level query requirements and the low-level facial features of the human facial image. Semantic features were selected and weighted based on a case study, with the participation of one hundred respondents. Visual features and semantic features were extracted by different methods, so they have variant weights. A new method was proposed through a radial basis function network for both, measuring the distance between the query vectors and the database vectors of the different features for similarities finding, and for ranking and combining the similarities. A probabilistic approach was used to improve the differences observed based on humans’ perception and the viewpoint that may appear during image annotation and/or query process. A prototype system of human facial image retrieval was subsequently built to test the retrieval performance. The system was trained and tested on two databases; the first database being the ‘ORL Database of Faces’ from AT&T Laboratories, Cambridge, while the second database consists of local facial images database of one hundred and fifty participants from the University of Malaya (UM), Kuala Lumpur,and some of their friends and families outside the UM. The results of the experiments show that, as compared to the content-based facial image retrieval technique, the proposed methods of SCBFIR achieve the best performance based on the number of semantic features used. The content-based facial image retrieval technique achieves 80.60% and 89.51% accuracy, while the SCBFIR achieves 97.85 % and 99.39% accuracy for the first and second database respectively within the top 10 retrieved facial images.

    Item Type: Thesis (PhD)
    Additional Information: Ph.D -- Fakulti Sains Komputer dan Teknologi Maklumat, Universiti Malaya 2013
    Uncontrolled Keywords: Content-based image retrieval; Image processing--Digital techniques; Information retrieval
    Subjects: Q Science > QA Mathematics > QA76 Computer software
    Divisions: Faculty of Computer Science & Information Technology
    Depositing User: Mrs Nur Aqilah Paing
    Date Deposited: 25 Jun 2015 23:06
    Last Modified: 25 Jun 2015 23:06
    URI: http://studentsrepo.um.edu.my/id/eprint/5724

    Actions (For repository staff only : Login required)

    View Item