Unsupervised object segmentation aims to assign same label to pixels of object region with feature homogeneity, which can be applied to object detection and recognition. Intersecting cortical model (ICM) can simulate human visual system (HVS) to process image for many applications, and at the same time, saliency detection can also simulate HVS to locate the most important object in a scene. Based on saliency detection, a novel approach for unsupervised object segmentation, termed as saliency guided intersecting cortical model (SG-ICM), is proposed in this paper. Instead of using gray-scale and spatial information to motivate ICM neurons traditionally, it is better to exploit saliency characteristic to guide ICM. In this paper, we plan to do saliency detection exploiting an improved dynamic guided filtering to analyze significance of different regions in same scene. The proposed saliency feature lies on: (1) the proposed saliency detection is based on region instead of pixel; (2) the dynamic guided filter is designed to accelerate the filtering; (3) in order to improve SG-ICM for object segmentation, at the each iteration, we use adaptive and simple threshold, which can raise the speed of this model. We check the proposed algorithm on common database of DOTI, color image from public database of MSRA with ground truth annotation. Experimental results show that the proposed method is superior to the others in terms of robustness of object segmentation, furthermore, it does not need any training. In addition, this method is effective for aerial image, the detection results reveal that this model has great potential in aerial reconnaissance application.
[1]
Weiqiang Wang,et al.
Saliency-guided Pairwise Matching
,
2017,
Pattern Recognit. Lett..
[2]
Nie Shou-ping.
Image Segmentation Algorithm Study for Low Contrast Image
,
2004
.
[3]
Hong-Bin Shen,et al.
Saliency driven region-edge-based top down level set evolution reveals the asynchronous focus in image segmentation
,
2018,
Pattern Recognit..
[4]
Ulas Bagci,et al.
Automatically Designing CNN Architectures for Medical Image Segmentation
,
2018,
MLMI@MICCAI.
[5]
Huchuan Lu,et al.
CNN for saliency detection with low-level feature integration
,
2017,
Neurocomputing.
[6]
Adrian G. Bors,et al.
Image retrieval based on query by saliency content
,
2015,
Digit. Signal Process..
[7]
Huchuan Lu,et al.
SaliencyRank: Two-stage manifold ranking for salient object detection
,
2015,
Computational Visual Media.
[8]
Reinhard Eckhorn,et al.
Feature Linking via Synchronization among Distributed Assemblies: Simulations of Results from Cat Visual Cortex
,
1990,
Neural Computation.