UPM Institutional Repository

Improved salient object detection via boundary components affinity


Nadzri, Nur Zulaikhah and Marhaban, Mohammad Hamiruce and Ahmad, Siti Anom and Ishak, Asnor Juraiza and Mohd Zin, Zalhan (2019) Improved salient object detection via boundary components affinity. Pertanika Journal of Science & Technology, 27 (4). pp. 1735-1758. ISSN 0128-7680; ESSN: 2231-8526


Referring to the existing model that considers the image boundary as the image background, the model is still not able to produce an optimum detection. This paper is introducing the combination features at the boundary known as boundary components affinity that is capable to produce an optimum measure on the image background. It consists of contrast, spatial location, force interaction and boundary ratio that contribute to a novel boundary connectivity measure. The integrated features are capable to produce clearer background with minimum unwanted foreground patches compared to the ground truth. The extracted boundary features are integrated as the boundary components affinity. These features were used for measuring the image background through its boundary connectivity to obtain the final salient object detection. Using the verified datasets, the performance of the proposed model was measured and compared with the 4 state-of-art models. In addition, the model performance was tested on the close contrast images. The detection performance was compared and analysed based on the precision, recall, true positive rate, false positive rate, F Measure and Mean Absolute Error (MAE). The model had successfully reduced the MAE by maximum of 9.4%.

Download File

17 JST-1475-2018.pdf

Download (2MB) | Preview

Additional Metadata

Item Type: Article
Divisions: Faculty of Engineering
Publisher: Universiti Putra Malaysia Press
Keywords: Boundary connectivity; Boundary ratio; Force interaction
Depositing User: Nabilah Mustapa
Date Deposited: 04 Feb 2020 03:50
Last Modified: 04 Feb 2020 03:50
URI: http://psasir.upm.edu.my/id/eprint/76320
Statistic Details: View Download Statistic

Actions (login required)

View Item View Item