We propose a simple yet powerful Boolean Map based Saliency (BMS) model for computing saliency maps that highlight visually salient regions in images. The method is inspired by the Gestalt principle of figure-ground segregation: it generates a set of Boolean maps by randomly thresholding the feature maps of an input image, and then computes the saliency map by analyzing the topological structure of these Boolean maps. Unlike most existing bottom-up saliency models that rely on center-surround contrast computation, our approach exploits global properties of the image to detect salient regions. Experimental results on five eye tracking benchmark datasets demonstrate that the proposed method consistently outperforms ten state-of-the-art saliency models. In addition, we show that BMS also performs well in salient object detection tasks.

Visual saliency detection has been a fundamental problem in computer vision and neuroscience. A saliency map is a topographic representation that highlights conspicuous regions in an image, and it plays an important role in many applications such as image retargeting, object recognition, and image quality assessment. Most existing bottom-up saliency models are based on the center-surround contrast hypothesis, which computes saliency by measuring the local contrast between a region and its surrounding context. While effective in many cases, this paradigm has inherent limitations in handling large salient objects or cluttered backgrounds.

In this paper, we draw inspiration from the Gestalt principle of figure-ground segregation. According to this principle, the human visual system tends to perceive enclosed regions as figures against the surrounding ground. We operationalize this idea through Boolean maps: binary images obtained by thresholding color channels of the input. Each Boolean map partitions the image into foreground (figure) and background (ground) regions. The saliency of a region is then related to its surroundedness — how much it is enclosed by other regions. By computing the mean of the attention maps derived from all Boolean maps, we obtain a robust saliency map that captures global figure-ground structure.

The seminal Itti-Koch model computes saliency using multi-scale center-surround differences across color, intensity, and orientation channels. Subsequent models such as GBVS (Graph-Based Visual Saliency) and AIM (Attention based on Information Maximization) improve upon this framework but remain fundamentally local in nature. Frequency domain approaches like the Spectral Residual model offer an alternative perspective by analyzing the log-spectrum of an image, yet they lack explicit connection to perceptual principles. More recent methods incorporate global context through techniques like Bayesian surprise or information-theoretic measures, but they often come with high computational costs. Our approach uniquely combines global structure analysis with computational simplicity.

Given an input image, we first decompose it into a set of feature channels including the L, a, b channels in CIELAB color space. For each feature channel, we generate multiple Boolean maps by applying a series of uniformly sampled thresholds. Specifically, for a feature channel with values normalized to [0, 1], we threshold at values t_1, t_2, ..., t_N where these thresholds are evenly spaced. Each threshold produces a binary image (Boolean map) where pixels above the threshold are set to 1 (foreground) and the rest to 0 (background). This process generates a comprehensive set of figure-ground hypotheses that collectively explore different possible segmentations of the image.

For each Boolean map, we compute an attention map based on the surroundedness of the foreground regions. The key insight is that a connected foreground region that is fully surrounded by background is more likely to be salient than one touching the image border. We implement this using morphological operations: specifically, we apply a flood fill operation from the image borders to identify background-connected regions, then the remaining foreground pixels constitute the "surrounded" regions. The attention value of each surrounded region is set proportional to its area. The final saliency map is obtained by averaging the attention maps across all Boolean maps and all feature channels, followed by Gaussian smoothing and post-processing normalization.

We evaluate BMS on five widely-used eye tracking benchmark datasets: MIT, Toronto, NUSEF, Kootstra, and SUN. The method is compared against ten state-of-the-art saliency models including Itti-Koch, GBVS, AIM, SUN, DVA, SIG, AWS, Murray, HouCVPR, and Judd. We use standard evaluation metrics including AUC (Area Under the ROC Curve), sAUC (shuffled AUC), and CC (Correlation Coefficient). Results show that BMS achieves the best or second-best performance across all five datasets under all three metrics. Notably, BMS achieves an average AUC of 0.933 on the MIT dataset, significantly outperforming the Itti-Koch baseline (0.879). Furthermore, we evaluate BMS on the MSRA-B salient object detection dataset, where it also demonstrates competitive performance against dedicated salient object detectors, confirming the generality of the approach.

We have presented BMS, a Boolean Map based Saliency model that detects visually salient regions by exploiting the topological structure of Boolean maps. Inspired by the Gestalt principle of figure-ground segregation, our method computes saliency based on the surroundedness of regions, offering a fundamentally different perspective from the dominant center-surround contrast paradigm. The method is straightforward to implement, computationally efficient, and achieves state-of-the-art performance across multiple benchmark datasets. We believe that the Boolean map representation provides a rich and underexplored framework for visual attention modeling, and future work could extend this approach to incorporate temporal information for video saliency detection.