A top-down region dividing approach for image segmentation

doi:10.1016/j.patcog.2007.11.020

Pattern Recognition

Volume 41, Issue 6, June 2008, Pages 1948-1960

https://doi.org/10.1016/j.patcog.2007.11.020 Get rights and content

Abstract

Histogram-based and region-based segmentation approaches have been widely used in image segmentation. Difficulties arise when we use these techniques, such as the selection of a proper threshold value for the histogram-based technique and the over-segmentation followed by the time-consuming merge processing for the region-based technique. To provide efficient algorithms that not only produce better segmentation results but also maintain low computational complexity, a novel top-down region dividing based approach is developed for image segmentation, which combines the advantages of both histogram-based and region-based approaches. Experimental results show that our algorithm can efficiently perform image segmentation without distorting the spatial structure of an image. Furthermore, two potential applications in medical image analysis are presented to show the advantages of using the proposed algorithm.

Introduction

Image segmentation aims to extract or threshold objects with respect to the background. It is a critical preprocessing step to the success of image recognition [1], image compression [2], image visualization [3], and image retrieval [4]. There are many types of image segmentation techniques [2], [3], [4], [5], [6], [7]. Among them, the histogram-based and the region-based algorithms are most popular.

The histogram-based (or feature-based) segmentation technique produces a binary image based on the threshold value [6]. The intensities of object and background pixels tend to cluster into two sets in the histogram. The histogram will be bimodal by a threshold value, which is selected from the valley between the two sets. The optimal segmentation intends to find the threshold value that minimizes the misclassification. If the threshold is too high, many object pixels will be lost and object contour will be severely destroyed. Although the complexity of histogram-based technique is low, the threshold selection is difficult, especially when the histogram is multi-modal. Furthermore, the histogram-based technique only considers the feature image (histogram) without checking the spatial relationship among connected pixels.

Watershed-based segmentation, a region-based approach, uses a bottom-up strategy that segments an image into several small regions, followed by a merge procedure. The immersion-based system [8] and the drainage rainfall system [9] are two approaches for performing watershed transformation. It considers an image as a topographic surface and the image intensity as the altitude. The drop of water will progressively fill up the ascending catchment basins from the minima of lowest altitude (lowest intensity) of the surface. Each pixel will flow along a descending path to a local minimum. When the altitude of water gradually increases, two catchment basins will reach at some points, called watershed points. A collection of watershed pixels on the contour is defined as the watershed line. Since the watershed algorithm is highly sensitive to the local minimum, it usually results in over-segmentation. In other words, there are overcrowded regions segmented in an image. Furthermore, in order to merge similarly smaller connected regions, the region adjacency graph (RAG) is used for region growing. Although the watershed-based image segmentation provides better results than the histogram-based approach, its computational complexity is high.

In order to provide more efficient algorithms that not only obtain better results, but also maintain low complexity, a novel top-down region dividing (TDRD) based approach is developed to iteratively divide sub-regions if the size of a sub-region is larger than a predefined threshold or the homogeneity of a sub-region is larger than a predefined threshold. The rest of this paper is organized as follows. In Section 2, we present the overview of the proposed TDRD-based image segmentation. The region dividing and sub-region examination strategies are provided in Section 3. Experimental results are shown in Section 4. Section 5 introduces its potential applications to medical image analysis. Finally, conclusions are drawn in Section 6.

Section snippets

Problem motivation

The histogram-based image segmentation method, although its complexity is low, does not consider the spatial relationship of neighborhood and may fail in some cases, such as the case in Fig. 1. Fig. 1(a) shows an image containing dark and bright parts on the left and right areas, respectively. In each part, gray values gradually increase from left to right. For example, the pixels on the leftmost column of left part are “32,” and are increased by 1 on each column to the right, until the pixels

Region dividing procedure

The region dividing procedure is based on our previous image simplification algorithm [10] that combines the advantages from histogram-based and region-based segmentation methods. It includes three steps: (a) suspicious intensities determination, (b) suspicious pixels determination, and (c) final intensity determination (FID).

(a) Suspicious intensities determination: The suspicious intensities are determined by comparing the histograms of two transformed images using histogram equalization

Experimental results

We perform our algorithm using two dividing iterations on an input image in Fig. 5(a). After obtaining Fig. 8(a) which performs one dividing iteration, we continuously segment all the sub-regions to achieve two dividing iterations for an input image. The results of performing the second iteration of our TDRD-based image simplification are shown in Fig. 9. Since there are three classes in white, gray and black in Fig. 8(a), the results of the second iteration are shown below for each class.

(1)
Figs.

Potential applications in medical image analysis

Medical image analysis is important since it provides assistance for medical doctors to find out the diseases inside the body without the surgery procedure [15]. The TDRD-based image segmentation provides useful applications due to the properties of medical images. Generally, medical images contain three major regions: background, soft tissue, and object. We present the ideas of two potential applications, breast boundary segmentation and lung segmentation. Further detailed implementation is

Conclusions

We have presented a TDRD-based image segmentation technique to combine the advantages of histogram-based and region-based approaches. The TDRD-based algorithm consists of two major procedures: region dividing and sub-region evaluation. In the region dividing procedure, the suspicious pixels are obtained from the suspicious intensities. In the sub-region evaluation procedure, the final intensities of suspicious pixels are determined by considering local spatial information. Experimental results

About the Author—YI-TA WU was born in Taipei, Taiwan. He received the B.S. degree in Physics from Tamkang University, Taipei, Taiwan, in 1995, and the M.S. degree in Computer Science from National Dong-Hwa University, Hualien, Taiwan, in 1997. Dr. Wu received Ph.D. from Department of Computer Science, New Jersey Institute of Technology, in May 2005. He is now a research fellow at University of Michigan, Ann Arbor. His current research interests include image/video processing, mathematical

References (25)

C.-S. Lee et al.
Weighted fuzzy mean filters for image processing
Fuzzy Sets and Systems
(1997)
W.-L. Hung et al.
Fuzzy clustering on LR-type fuzzy numbers with an application in Taiwanese tea evaluation
Fuzzy Sets and Systems
(2005)
U. Bick et al.
Automated segmentation of digitized mammograms
Acad. Radiol.
(1995)
P.W. Pachowicz
Semi-autonomous evolution of object models for adaptive object recognition
IEEE Trans. System Man Cybern.
(1994)
K. Belloulata et al.
Fractal image compression with region-based functionality
IEEE Trans. Image Process.
(2002)
S.L. Hartmann et al.
Depth-buffer targeting for spatially accurate 3-D visualization of medical images
IEEE Trans. Medical Imaging
(2000)
Y. Chen et al.
A region-based fuzzy feature matching approach to content-based image retrieval
IEEE Trans. Pattern Anal. Machine Intell.
(2002)
K. Haris et al.
Hybrid image segmentation using watersheds and fast region merging
IEEE Trans. Image Process.
(1998)
O.J. Tobias et al.
Image segmentation by histogram thresholding using fuzzy sets
IEEE Trans. Image Process.
(2002)
A. Khotanzad et al.
Unsupervised segmentation of textured images by edge detection in multidimensional feature
IEEE Trans. Pattern Anal. Machine Intell.
(1989)

L. Vincent et al.

Watersheds in digital spaces: an efficient algorithm based on immersion simulations

IEEE Trans. Pattern Anal. Machine Intell.

(1991)

R. Chibanga, J. Berlamont, J. Vandewalle, Artificial neural networks in hydrological watershed modeling: surface flow...

Cited by (32)

Lung nodule segmentation using Salp Shuffled Shepherd Optimization Algorithm-based Generative Adversarial Network
2021, Computers in Biology and Medicine
Citation Excerpt :
Recent researchers on lung CT image segmentation focussed on accurate and fast segmentation methods. The methods for the pulmonary nodule segmentation are grouped as thresholding method, morphological method, deformable model, clustering method, neural networks, graph cut method, active contours, watershed, histogram [9], fuzzy logic-based segmentation, region growing [10], and Markov random field. Moreover, the thresholding method is one of the traditional methods used for segmentation [11].
Lung nodule segmentation is an exciting area of research for the effective detection of lung cancer. One of the significant challenges in detecting lung cancer is Accuracy, which is affected due to the visual deviations and heterogeneity in the lung nodules. Hence, to improve the segmentation process's Accuracy, a Salp Shuffled Shepherd Optimization Algorithm-based Generative Adversarial Network (SSSOA-based GAN) model is developed in this research for lung nodule segmentation. The SSSOA is the hybrid optimization algorithm developed by integrating the Salp Swarm Algorithm (SSA) and shuffled shepherd optimization algorithm (SSOA). The artefacts in the input Computed Tomography (CT) image are removed by performing pre-processing with the help of a Gaussian filter. The pre-processed image is subjected to lung lobe segmentation, which is done with the help of deep joint segmentation for segmenting the appropriate regions. The lung nodule segmentation is performed using the GAN. The GAN is trained using the SSSOA to effectively segment the lung nodule from the lung lobe image. The metrics, such as Dice Coefficient, Accuracy, and Jaccard Similarity, are used to evaluate the performance. The developed SSSOA-based GAN method obtained a maximum Accuracy of 0.9387, a maximum Dice Coefficient of 0.7986, and a maximum Jaccard Similarity of 0.8026, respectively, compared with the existing lung nodule segmentation method.
Saliency detection based on directional patches extraction and principal local color contrast
2018, Journal of Visual Communication and Image Representation
Citation Excerpt :
Oliva et al. [38] proposed a top-down saliency detection framework based on global scene configuration for object detection. In [39], a top-down computational mechanism with prior knowledge and region dividing strategy was devised for image segmentation. Cholakkal et al. [40] designed a top-down saliency detection approach via locality-constrained contextual sparse coding.
Saliency detection has become an active topic in both computer vision and multimedia fields. In this paper, we propose a novel computational model for saliency detection by integrating the holistic center-directional map with the principal local color contrast (PLCC) map. In the proposed framework, perceptual directional patches are firstly detected based on discrete wavelet frame transform (DWFT) and sparsity criterion, then the center of the spatial distribution of the extracted directional patches are utilized to locate the salient object in an image. Meanwhile, we proposed an efficient local color contrast method, called principal local color contrast (PLCC), to compute the color contrast between the salient object and the image background, which is sufficient to highlight and separate salient objects from complex background while dramatically reduce the computational cost. Finally, by incorporating the complementary visual cues of the global center-directional map with the PLCC map, a final compounded saliency map can be generated. Extensive experiments performed on three publicly available image databases, verify that the proposed scheme is able to achieve satisfactory results compared to other state-of-the-art saliency-detection algorithms.
An inverse approach for automatic segmentation of carotid and vertebral arteries in CTA
2018, Expert Systems with Applications
Carotid artery stenosis is generally a constriction caused by atherosclerosis or carotid artery lumen bottleneck. Carotid arteries are located closely to bones and osteoid structures. Osteoid structures and carotid arteries are frequently confused with each other when performing vessel evaluations. This study provides a novel method for carotid artery lumen segmentation on CTA images by using automatic vessel segmentation with inverse approach, in which vessel segmentation is performed after bone region is segmented and eliminated. The region growing and random walk segmentation methods are utilized in the elimination of bone region and the vessel segmentation. The seed points in the mentioned methods are not manually determined by any starting point. In automatic segmentation, seeds are selected from the experimentally determined intervals according to the local histogram. The stages of preprocessing and post-processing are utilized for better segmentation. The tracking of vessel centers based on continuity is employed for 3D reconstruction and 3D imaging of the vessels. Experiments were conducted with different data sets including various CTA images by using the mentioned methods. As a result, dice similarity rate above 92% was achieved together with 0.16 mm Msd and 99% accuracy. It was concluded based on these results that the proposed method provides successful results in different points of common, internal, external, vertebral arteries, carotid bifurcation and locales close to osteoid structures which are deemed challenging regions for carotid artery lumen segmentation.
Segmentation of lung nodule in CT data using active contour model and Fuzzy C-mean clustering
2016, Alexandria Engineering Journal
Citation Excerpt :
It partitions the image into segments corresponding to the anatomical objects in the image. In the recent years, a lot of pulmonary nodule segmentation methods have been proposed, which can be categorized as thresholding method [7,8], morphological method [9], deformable model [10], clustering method [11–13], graph cut method [14,15], Markov random field, region growing [16], watershed, neural networks, fuzzy logic, active contours [17] and histogram based segmentation [18]. Among various segmentation methods, active contour was one of the most popular and successful one.
The aim of this paper was to develop a region based active contour model and Fuzzy C-Means (FCM) technique for segmentation of lung nodules. Ultimately, detection and assisted diagnosis of nodules at earlier stage increase the mortality rate. Among many imaging modalities, Computed Tomography (CT) is being the most sought because of its imaging sensitivity, high resolution and isotropic acquisition in locating the lung lesions. The proposed methodology focuses on acquisition of CT images, reconstruction of lung parenchyma and segmentation of lung nodules. Reconstruction of parenchyma can be employed using selective binary and Gaussian filtering with new signed pressure force function (SBGF-new SPF) and clustering technique was used for nodule segmentation. Comparative experiments demonstrate the advantages of the proposed method in terms of decreased error rate and increased similarity measure.
Oriented boundary graph: An efficient structuring model for segmentation of 3D images
2016, Computer Vision and Image Understanding
Citation Excerpt :
The segmentation process aims to define a partition of the image into relevant regions according to given criteria. Different approaches exist to achieve this segmentation like split and merge methods [1–5], which are based on regions, active contour methods [6–9] based on region boundaries, and Markov methods [10,11] using a probabilistic approach. Split and merge methods consists in alternatively splitting and merging the regions of the partition.
From a theoretical point of view, most of image segmentation methods that have been developed for 2D images can be generalized to higher dimensions. In actual practice, the cost in space to encode 3D data structure and the cost in time to run 3D algorithms does not allow to conveniently implement those classical segmentation algorithms in the 3D case. In this article, we describe a new model to efficiently represent and update both the topological and the geometrical structure of the regions of a 3D segmented image. This model has been defined from a pragmatical approach that consists in specifying a basic region-based segmentation framework, and then in building a minimal model that encodes all the relationships needed for an efficient implementation of this framework. This approach leads to a model suitable for a wide range of segmentation methods and allowing an efficient computation of most of the segmentation criteria involved in image segmentation.
Bayesian salient object detection based on saliency driven clustering
2014, Signal Processing: Image Communication
Citation Excerpt :
Visual attention has been studied by researchers in physiology, psychology, neural systems, and computer vision for a long time. Extracting objects from an image is a hot research topic and has wide applications, such as content-based image retrieval [1], image/video compression and coding [2], object recognition and scene understanding [3–6] and image segmentation [7,8] in areas of computer vision and computer graphics. Under the mechanism of visual attention, HVS picks out relevant parts of a scene as attention regions corresponding to salient regions in images.
Salient object detection is essential for applications, such as image classification, object recognition and image retrieval. In this paper, we design a new approach to detect salient objects from an image by describing what does salient objects and backgrounds look like using statistic of the image. First, we introduce a saliency driven clustering method to reveal distinct visual patterns of images by generating image clusters. The Gaussian Mixture Model (GMM) is applied to represent the statistic of each cluster, which is used to compute the color spatial distribution. Second, three kinds of regional saliency measures, i.e, regional color contrast saliency, regional boundary prior saliency and regional color spatial distribution, are computed and combined. Then, a region selection strategy integrating color contrast prior, boundary prior and visual patterns information of images is presented. The pixels of an image are divided into either potential salient region or background region adaptively based on the combined regional saliency measures. Finally, a Bayesian framework is employed to compute the saliency value for each pixel taking the regional saliency values as priority. Our approach has been extensively evaluated on two popular image databases. Experimental results show that our approach can achieve considerable performance improvement in terms of commonly adopted performance measures in salient object detection.

View all citing articles on Scopus

About the Author—FRANK Y. SHIH received the B.S. degree from National Cheng-Kung University, Taiwan, in 1980, the M.S. degree from the State University of New York at Stony Brook, in 1984, and the Ph.D. from Purdue University, West Lafayette, Indiana, in 1987, all in Electrical Engineering. He is presently a professor jointly appointed in the Department of Computer Science, the Department of Electrical and Computer Engineering, and the Department of Biomedical Engineering at New Jersey Institute of Technology, Newark, NJ. He currently serves as the Director of Computer Vision Laboratory.

Dr. Shih is currently on the Editorial Board of the International Journal of Pattern Recognition, the International Journal of Pattern Recognition Letters, the International Journal of Pattern Recognition and Artificial Intelligence, the International Journal of Recent Patents on Engineering, the International Journal of Recent Patents on Computer Science, the International Journal of Internet Protocol Technology, and the Journal of Internet Technology. Dr. Shih has contributed as a steering member, committee member, and session chair for numerous professional conferences and workshops. He was the recipient of the Research Initiation Award from the National Science Foundation in 1991. He won the Honorable Mention Award from the International Pattern Recognition Society for Outstanding Paper and also won the Best Paper Award in the International Symposium on Multimedia Information Processing. He has received several awards for distinguished research at New Jersey Institute of Technology. He has served several times on the Proposal Review Panel of the National Science Foundation.

Dr. Shih holds the research fellow for the American Biographical Institute and the IEEE senior membership. He has published seven book chapters and over 180 technical papers, including 85 in well-known prestigious journals. His current research interests include image processing, computer vision, sensor networks, pattern recognition, bioinformatics, information security, robotics, fuzzy logic, and neural networks.

About the Author—JIAZHENG SHI received B.E. and M.E. from Beijing University of Posts & Telecom, respectively, in 1997 and 2000, and received Ph.D. from University of Nebraska-Lincoln in 2005. He is now a research fellow at University of Michigan, Ann Arbor. His research interests include image processing and medical imaging analysis.

View full text

A top-down region dividing approach for image segmentation

Abstract

Introduction

Section snippets

Problem motivation

Region dividing procedure

Experimental results

Potential applications in medical image analysis

Conclusions

Fuzzy Sets and Systems

Fuzzy Sets and Systems

Acad. Radiol.

Semi-autonomous evolution of object models for adaptive object recognition

IEEE Trans. System Man Cybern.

Fractal image compression with region-based functionality

IEEE Trans. Image Process.

Depth-buffer targeting for spatially accurate 3-D visualization of medical images

IEEE Trans. Medical Imaging

A region-based fuzzy feature matching approach to content-based image retrieval

IEEE Trans. Pattern Anal. Machine Intell.

Hybrid image segmentation using watersheds and fast region merging

IEEE Trans. Image Process.

Image segmentation by histogram thresholding using fuzzy sets

IEEE Trans. Image Process.

Unsupervised segmentation of textured images by edge detection in multidimensional feature

IEEE Trans. Pattern Anal. Machine Intell.

Watersheds in digital spaces: an efficient algorithm based on immersion simulations

IEEE Trans. Pattern Anal. Machine Intell.