Unsupervised image segmentation based on analysis of binary partition tree for salient object extraction
Introduction
Salient object extraction from images and videos is usually an important part in many multimedia applications such as object-based coding, object-based image/video retrieval, image/video editing and manipulation, smart video surveillance, and human computer interaction. As a preceding step, a suitable image segmentation result will greatly facilitate the following process of salient object extraction. In view of the requirement for salient object extraction, the most preferred image segmentation result should be represented by possibly fewer segmented regions that can still preserve the boundaries of salient objects well. A number of traditional image segmentation approaches such as [1], [2], [3], [4], [5], [6] are generally exploited to obtain a region segmentation result, in which pixels in each segmented region share similar intensity, color or texture, but the problems of over-segmentation and under-segmentation are usually unavoidable. Furthermore, it is possible to obtain a hierarchical segmentation result with different number of segmented regions by directly adjusting some parameters in the so-called multi-scale (multi-resolution) segmentation approaches [7], [8], [9], but it requires a set of manually tuned parameters to obtain a suitable segmentation result for salient object extraction.
Binary partition tree (BPT) was introduced in [10] to systematically represent the hierarchical segmentation of an image in an efficient way. Starting from an initial over-segmentation result generated by any image segmentation approach, the simple yet effective region merging scheme can be exploited to progressively merge adjacent regions based on some kind of dissimilarity measures. The merging sequence can be efficiently recorded by BPT, in which each leaf node represents each initially segmented region and each non-leaf node represents the newly generated region during the region merging process. By using nodes at different levels in BPT to represent the image, it is convenient to obtain a segmentation result at any scale (with any number of segmented regions). By incorporating prior knowledge of a specific class of objects, automatic extraction of face objects [11], [12] and moving objects [13] can be realized by analysis on individual node in the BPT. For general salient object extraction without any prior knowledge, BPT analysis is also useful for highlighting salient regions in the image. In previous works, BPT simplification based on evolvement of region statistics is proposed for convenient tree visualization [14], and is used for efficient image segmentation and interactive extraction of salient objects [15].
In this paper, we present a novel BPT analysis work for unsupervised image segmentation, which shows the suitability for the application of salient object extraction. From an over-segmentation result and the generated BPT, the proposed BPT analysis algorithm automatically selects an appropriate subset of nodes to represent a more meaningful segmentation result. Compared with the previous works [14], [15] based on BPT analysis, the main contribution of our work is twofold. One is that a novel dissimilarity measure considering the impact of color difference, area factor and adjacency degree in a unified way is proposed for region merging and used in the BPT generation process. The other is the proposed BPT analysis algorithm, in which the node evaluation is designed to reasonably identify salient regions, and the following two-phase node selecting process guarantees a meaningful segmentation result possibly reserving salient regions. An obvious feature of our approach is totally free of threshold, while the previous works [14], [15] need user-supplied thresholds during the BPT analysis process. As an unsupervised image segmentation approach, our approach improves the segmentation performance from the view of salient object extraction.
The remainder of this paper is organized as follows. Section 2 describes the process of BPT generation with region merging from an initial segmentation. Section 3 details the proposed BPT analysis algorithm for a meaningful segmentation. Experimental results are shown in Section 4, and conclusion is given in Section 5.
Section snippets
BPT generation from initial segmentation
The original image can be initially partitioned into a set of homogenous regions using a collection of existing image segmentation approaches. The only issue when using any image segmentation approach and possibly adjusting its parameters is to avoid under-segmentation. In other words, the only requirement for initial segmentation is that each segmented region should possibly not cover the parts from different salient objects and background. In this paper, watershed transform [1] is exploited
BPT analysis for meaningful segmentation
In this section, we propose a systematic BPT analysis algorithm to select a suitable subset of nodes to represent a more meaningful segmentation. The proposed algorithm consists of the following two stages, that is, BPT node evaluation and BPT node selection, which are detailed in the following two subsections, respectively.
Experimental results
In order to evaluate the performance of the proposed BPT analysis based unsupervised image segmentation approach from the view of salient object extraction, we select 160 images containing at least one obvious salient object from Berkeley segmentation dataset (BSD) [16], Corel photo gallery, and our image collection. Experimental results on four representative test images are shown in Fig. 3, in which original images, initial segmentation results, final segmentation results generated using our
Conclusion
In this paper, we have presented an efficient unsupervised image segmentation approach based on the BPT analysis. From an initial segmentation result of the original image, a BPT is generated with the region merging process, which is controlled by a novel dissimilarity measure considering the impact of color difference, area factor, and adjacency degree in a unified way. By a systematic analysis of the evaluated BPT, a more meaningful segmentation result is represented by a small subset of
Acknowledgments
The authors are grateful to the anonymous reviewers and the handling editor for their valuable comments, which have greatly helped us to make improvements. This work is supported by National Natural Science Foundation of China under Grant No. 60602012, Shanghai Educational Development Foundation under Grant No. 2007CG53, and Innovation Program of Shanghai Municipal Education Commission (No. 09YZ02).
References (17)
- et al.
Image segmentation with a fuzzy clustering algorithm based on ant-tree
Signal Process.
(2008) - et al.
Scalable multiresolution color image segmentation
Signal Process.
(2006) - et al.
Face segmentation and tracking based on connected operators and partition projection
Pattern Recognition
(2002) - et al.
An efficient face segmentation algorithm based on binary partition tree
Signal Process.: Image Commun.
(2005) - et al.
Watersheds in digital spaces: an efficient algorithm based on immersion simulations
IEEE Trans. Pattern Anal. Machine Intell.
(1991) - et al.
Hybrid image segmentation using watersheds and fast region merging
IEEE Trans. Image Process.
(1998) - et al.
Statistical region merging
IEEE Trans. Pattern Anal. Machine Intell.
(2004) - et al.
Normalized cuts and image segmentation
IEEE Trans. Pattern Anal. Machine Intell.
(2000)
Cited by (22)
Content-aware image resizing: An improved and shadow-preserving seam carving method
2019, Signal ProcessingCitation Excerpt :Saliency detection approaches are categorized into two groups: bottom-up methods and top-down methods. The bottom-up methods are based on low-level features such as color, intensity, and orientation [26–30]; while top-down methods employ semantic information such as face and text [31]. In the proposed method inspired by the idea given in [32], for extracting the co-saliency map from the multiple images, we introduce a simple yet efficient bottom-up saliency detection method.
Towards real-time crops surveillance for disease classification: exploiting parallelism in computer vision
2017, Computers and Electrical EngineeringA fast region segmentation algorithm on compressed gray images using Non-symmetry and Anti-packing Model and Extended Shading representation
2016, Journal of Visual Communication and Image RepresentationCitation Excerpt :An efficient representation method can not only save the storage space of images, but also can reduce the time required for some image manipulations [1–3].
Use of Binary Partition Tree and energy minimization for object-based classification of urban land cover
2015, ISPRS Journal of Photogrammetry and Remote SensingData field-based transition region extraction and thresholding
2012, Optics and Lasers in EngineeringSaliency-directed color image segmentation using modified particle swarm optimization
2012, Signal ProcessingCitation Excerpt :The watershed algorithm [14], a typical region-based segmentation technique, often produces a lot of small but homogeneous regions, which need some merging operations to reduce the number of regions [15,16]. Liu et al. [16] proposed an unsupervised image segmentation approach. Starting from an over-segmentation color image, region merging is performed using a dissimilarity measure considering color difference, area factor, and adjacency degree, and a binary partition tree (BPT) is generated to record the whole region-merging sequence.