Object‐based attention: saliency detection using contrast via background prototypes

Abstract

An object-based attention model to predict visual saliency using contrast against the 'background prototypes' is presented. The proposed model automatically identifies a series of regions far away from the image centre as background prototypes. The visual saliency is then calculated using the colour contrast against these background prototypes. Promising experimental results demonstrate the effectiveness of the proposed model in terms of detection accuracy and implementation efficiency.Introduction: The human attention system (HAS) is able to quickly detect the most interesting regions in a given scene. An important and open issue in computer vision is to simulate the HAS in humanmachine interactions, which enables the hardware systems to automatically focus on and capture interesting target objects. Owing to the limited computational resources (less powerful processor and limited memory) of electronic equipments, the designed saliency models are expected to be easy to implement, while maintaining accurate and robust detection results. Therefore, considerable effort has been devoted to detecting salient regions over the past few years [1][2][3][4][5][6][7].Existing saliency models could be categorised into two classes: top-down [8-10] and bottom-up [6,[11][12][13]. Top-down methods employ high-level cues (e.g. face); however, it is hardly generalised since their learning process often needs numerous computing resources. On the contrary, bottom-up approaches mainly estimate the foreground saliency based on simple low-level image features (e.g. luminance, colour and orientation), thus they are more convenient to apply to the practical scenarios of human-machine interactions. As a pioneer work, Itti et al.[11] introduced a biologically inspired saliency model based on the centre-surround operation. Graph-based saliency models [6,12] are suggested to predict saliency following the principle of Markov random work theory. Some researchers attempted to detect irregularities such as visual saliency in the frequency domain [1,13]. These saliency models, however, have limited capacity in complex scenes, where the salient and background regions are often heterogeneous.Another instance of biological evidence shows that human visual attention is often attracted to the image centre [14], since the background tends to be located in the image boundary. Inspired by this biological fact, we propose a bottom-up saliency model relying on the contrast against 'background prototypes'. Here, 'background prototypes' mean the superpixels are located far away from the image centre, even in the boundary of an image. Another important aspect of saliency modelling is how to computationally measure visual saliency. The traditional models often estimate saliency using the centre-surround contrast [11] or the global contrast against the entire scene [10, 15]. On the contrary, we resort to calculating colour-based contrast with respect to the background prototypes. The experimental results show the effectiveness of our model in terms of more rob...

Multi Information Fusion Network for Saliency Quality Assessment

“…Extensive experiments on two publicly available databases such as DUT-OMRON [5] and ECSSD [6], show that by incorporating image content information, the proposed model leads to significantly improved quality prediction accuracy. Furthermore, we also validate the applicability of the proposed method in selecting optimal saliency map from a set of candidates for improving the performance of saliency detection [7], [8] and segmentation [3], [9]. The results demonstrate that the optimal selection by our method can significantly outperforms the best saliency detection al-Copyright c 2019 The Institute of Electronics, Information and Communication Engineers gorithms and existing quality assessment methods for both the saliency detection and saliency segmentation.…”

Automatic Salient Object Extraction Based on Locally Adaptive Thresholding to Generate Tactile Graphics

“…Cheng et al [17] demonstrated two saliency models: histogram-based contrast (HC), which assigns a pixel-wise saliency value, as well as region-based contrast (RC), which incorporates spatial relations at the cost of reduced computational efficiency. Zhou [25] proposed an object-based attention model that automatically identifies a series of regions far away from the image center as background prototypes. Zhou et al [26] combined widely used contrast measurements, namely, center-surround, corner-surround, and global contrast, to detect visual saliency.…”

Visual saliency object detection using sparse learning

“…In the bottom‐top saliency methods, low‐level features are used to extract the saliency map [14, 15]. These features encompass colour, texture, brightness levels, etc.…”