UPM Institutional Repository

Predicting saliency existence using reduced salient features based on compactness and boundary cues


Citation

Nadzri, Nur Zulaikhah (2020) Predicting saliency existence using reduced salient features based on compactness and boundary cues. Doctoral thesis, Universiti Putra Malaysia.

Abstract

Salient object detection is a process that tries to locate the most prominent region within the visual scene. Despite of hundreds of successful developed models, most of the models still fail to produce correct detection for background only image and this is due to the models’ assumption that a salient object must at least exist in an image. As for the previous models for saliency prediction, the developed models require complex computation for feature extraction with large number of salient features. Therefore, there is a need to develop a model for predicting saliency existence as it is able to categorise the input images as salient or non-salient. This research presented a method that can anticipate the existence of a salient object in the input image retrieved from ASD and SOSB datasets based on the compactness and boundary cues. This basis of assumption was adopted given the fact that the compactness of a non-salient image is spatially distributed as referred to its boundary compared to the salient image. The image background is measured based on the boundary contrast and boundary spatial distribution within the spatial and the frequency domain. These measurements represent the salient features that were extracted based on statistical computations that act as the global saliency identification. In order to reduce the number of extracted features, the salient features were selected based on filter method through the saliency histogram. The selected salient features were trained, tested and compared on 3 learning algorithms which included generalised linear regression, Naïve Bayes, and Support Vector Machine. By using the proposed method, the accuracy of the prediction was achieved where a maximum of 91.5% was obtained while other measurements of specificity, precision and recall obtained more than 90% accuracy. The validation of the developed model was tested on the current salient object detection model to observe the performance implication detecting the salient object. The saliency existence prediction on the integrated model achieved 92.5% for accuracy with the sacrifice of execution time which increased by 12.3% as compared to the salient object detection model BCA. This research does not only contribute to the model development but can also become an initial study focusing on saliency existence prediction that may have an impact on further model improvements in the future.


Download File

[img] Text
FK 2021 29 UPMIR.pdf

Download (1MB)

Additional Metadata

Item Type: Thesis (Doctoral)
Subject: Visual texture recognition
Subject: Computer science
Subject: Computer networks
Call Number: FK 2021 29
Chairman Supervisor: Professor Mohammad Hamiruce Marhaban, PhD
Divisions: Faculty of Engineering
Depositing User: Editor
Date Deposited: 13 Jul 2022 06:28
Last Modified: 13 Jul 2022 06:28
URI: http://psasir.upm.edu.my/id/eprint/98049
Statistic Details: View Download Statistic

Actions (login required)

View Item View Item