Salient Region Detection Using Patch Level and Region Level Image Abstractions

Abstract

Abstract-In this letter, a novel salient region detection approach is proposed. Firstly, color contrast cue and color distribution cue are computed by exploiting patch level and region level image abstractions in a unified way, where these two cues are fused to compute an initial saliency map. A simple and computationally efficient adaptive saliency refinement approach is applied to suppress saliency of background noises, and to emphasize saliency of objects uniformly. Finally, the saliency map is computed by integrating the refined saliency map with center prior map. In order to compensate different needs in speed/accuracy tradeoff, three variants of the proposed approach are also presented in this letter. The experimental results on a large image dataset show that the proposed approach achieve the best performance over several state-of-the-art approaches.Index Terms-saliency detection, color contrast, color distribution, center prior, adaptive saliency refinement. I. INTRODUCTIONDetecting salient regions in images is an interesting and difficult multidisciplinary problem. The field has considerable attention in the recent years, and has become an active area of research in Computer Vision due to its various applications in object detection, object recognition, adaptive image and video compression, and image retargeting. Many computational saliency detection models have been proposed over the years, which can be roughly categorized into bottom-up and topdown approaches [1].Bottom-up saliency is data-driven and is often estimated using color contrast cue Since color components of a salient object are always spatially compact rather than widely spread around the image, lower spatial distribution of a color component indicates its higher spatial saliency. Apart from these two cues, another widely used cue is center prior [1]. The center prior gives more

Discovering salient objects from videos using spatiotemporal salient region detection

“…For the comparative evaluation of our Patch-Region-based spatial salient region detection (PR) on a salient object detection image dataset MSRA-1000 [38], the readers are recommended to refer our work in [27]. Here, the robustness of the spatial salient region detection method to different numbers of patches and different numbers of regions, and the robustness of the faster variants were presented, as also were the performance of the individual saliency cues, the effect of spatial saliency refinement parameter and time comparison with different methods were also presented in our paper [27].…”

“…Here, the robustness of the spatial salient region detection method to different numbers of patches and different numbers of regions, and the robustness of the faster variants were presented, as also were the performance of the individual saliency cues, the effect of spatial saliency refinement parameter and time comparison with different methods were also presented in our paper [27].…”

“…Based on our previous work in [27], the spatial salient region detection approach presented in this paper combines both patch level and region level image abstractions which were separately considered for salient region detection in many others' works. Furthermore, a computationally efficient saliency refinement approach is presented to solve the saliency assignment issues in patch level saliency detection.…”

Improving Robustness for Vision Transformer with a Simple Dynamic Scanning Augmentation

“…This drop in performance is attributed by Dosovitskiy et al (2020) to the lack of inductive biases, such as locality, which are inherent to CNNs. However, studies from Lu et al (2016Lu et al ( , 2019; Kannan et al (2014) shows that the patch-matching algorithm can also be invariant to rotation solving a particular inductive bias. This suggests that performance in transformers can be further improved with better feature matchings.…”

Region-of-Interest Extraction Based on Local–Global Contrast Analysis and Intra-Spectrum Information Distribution Estimation for Remote Sensing Images

“…Since a salient target is usually composed of spatially connected salient pixels, a pixel surrounded by highly salient pixels is likely to be a part of the salient targets. On the other hand, a pixel enclosed by lower salient pixels is likely to be a part of background [46]. Therefore, we design a pixel-level saliency enhancement operation for S as follows:…”