Scene recognition with objectness. (February 2018)

• Record Type: Journal Article
• Authors: Cheng, Xiaojuan; Lu, Jiwen; Feng, Jianjiang; Yuan, Bo; Zhou, Jie
• Is Part Of: Pattern recognition. Volume 74(2018:Feb.)
• Page Start: 474
• Page End: 487
• Publication Date: 2018-02
• Subjects: Scene recognition -- Deep learning -- Co-occurrence pattern; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.09.025
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We exploit the correlations of object configurations among different scenes to choose discriminative objects and represent image descriptors with the occurrence probabilities of discriminative objects, which eliminate the negative effects caused by common objects to enhance the inter-class discriminability. We employ a new method patch screening to prune the patches containing non-discriminative objects by the intersection of top scored objects in patches and the discriminative objects, so that we improves the generalized characteristics of the same scenes. We validate the usefulness of our SDO with the state-of-the-art performance on three benchmark datasets: Scene 15, MIT Indoor 67 and SUN 397 datasets. Abstract: In this paper, we present a feature description method called semantic descriptor with objectness (SDO) for scene recognition. Most existing scene representation methods exploit the characteristics of constituent objects in scenes with inter-class independence, which ignore the negative effects caused by the common objects among different scenes. The generic characteristics of the common objects cause some generality among different scenes, which weakens the discriminative characteristics among scenes. To address this problem, we exploit the correlations of object configurations among different scenes by the co-occurrence pattern of all objects across scenes to choose representative and discriminative objects which enhances the inter-class discriminability. Specifically, we capture the statistic information of objects appearing in each scene to compute the distribution of each object across scenes, which obtains the co-occurrence pattern of objects. Moreover, we represent the image descriptors with the occurrence probabilities of discriminative objects in image patches to eliminate the negative effects of common objects. To make image descriptors more discriminative, we discard the patches with non-discriminative objects to enhance the intra-class generalized characteristics. Experimental results on three widely used scene recognition datasets show that our method outperforms the state-of-the-art methods.