A multi-level thresholding image segmentation based on an improved artificial bee colony algorithm☆
Introduction
Image segmentation technique is attempts to detect specific parts in an image, which is an important component in image processing, video processing and analysis [1], [2], [3]. As an important branch of image segmentation algorithm, thresholding methods segment a digital image into multiple parts. According to the number of thresholding, they have been divided into two categories: bi-level and multi-level algorithms. The former means an image should be divided into two subdivisions which use one grey value to represent its threshold. The multi-level method discriminates several distinct subdivisions from a digital image with more than one threshold. As a representative threshold-based segmentation method, Otsu [4] has attracted many researchers to do further study. Based on previous findings, the Otsu method can be treated as a maximum optimization problem. But a traditional exhausted searching method expends too much computational time to endure, especially on multi-level threshold selection of Otsu [5], [6].
As a population-based algorithm, evolutionary algorithms (EAs), e.g., differential algorithm (DE) and particle swarm optimization (PSO), find a potential solution space by employing multiple individuals, which means they could achieve fast computational ability than the traditional exhausted searching methods [7], [8], [9], [10], [11], [12]. As an efficient EA, artificial bee colony (ABC) has the characteristics of strong global search ability and steady robustness. Compared with other EAs, ABC shows a specific characteristic, i.e., one-dimension search strategy, which means bees in ABCs search the potential solutions one by one dimension. Then the global search ability of ABC is guaranteed but it should slow down its convergence rate [13], [14]. In this paper, by conquering the shortcomings of the ABC algorithm, we put forward an improved ABC algorithm to solve image segmentation problem successfully.
The remainder of this paper is organized as follows. Section 2 introduces the traditional ABC algorithm in detail. Then we propose a new ABC algorithm in section 3. Otsu for multi-level thresholding image segmentation is presented in Section 4. In Section 5, experiments on Berkeley image database is conducted and an integrate conclusion is drawn in Section 6.
Section snippets
Classical ABC algorithm
ABC algorithm was first proposed by Karaboga in 2005 [15], and its essential principle is to simulate the bee's foraging behavior, with utilizes a large amount of nectar to find food sources. The algorithm consists of three types of bees, i.e., employed bees, onlooker bees and scout bees, each of which corresponds to a different search task. The main flow can be described as follows.
- I.
Initialization: At first, all the bees are randomly initialized in a potential solution space. The maximum and
Improved algorithm
As descripted in the second sections, ABC employs a one-dimensional update strategy, which equips the algorithm with a strong global search capability but leads to two shortcomings: unsatisfied local search ability and slow convergence speed. Specifically, the update strategy of ABC algorithm should randomly choose one dimension to learn from a selected neighbor. It may result in stagnant for individual that selects the neighbors with bad performance. Then it should waste function evaluations
Otsu segmentation
Suppose there is a minimum gray level 0 and a maximum gray level L in an image which has N pixels. For ith grey level, the number of pixels is h(i) and the probability denoted as PRi.
The Otsu segmentation method should discriminate M − 1 thresholds, {th1,th2,..., thM − 1} of an image. Let Cj(j = 1, 2, …M) represent the ith part of grey level. Then we have C1 for [0, ⋅⋅⋅th1], ……, CM for [thM + 1, ⋅⋅⋅L]. The Otsu method determines the optimal thresholds by using the
Experimental results and comparison
Fig. 3 shows the original tested images of two selected Berkeley Image Database and their corresponding histogram of the gray scale. We compare the segmentation results with nine different algorithms, i.e., DE [19], PSO [20], QPSO [21], ABC [15], gABC [16], I_ABC [22], QABC [23], OLABC [24] and FGABC, on 31 standard test images [25]. Tables 1 and 2 show the mean values of the first six images and the total ranking of the nine algorithms on the 31 standard test images respectively. From the
Conclusion
We put forward an ABC algorithm with an improved search strategy with faster convergence speed. Based on leading by two promising points, our algorithm could search the area defined by the points more precisely, which should accelerate the convergence rate of population. Furthermore, equipping with an adaptive parameter, the new algorithm has more chances to search in the local area of the best points. Experimentations on benchmark images have demonstrated its efficiency and effectiveness. Our
Acknowledgment
The authors acknowledge support from National Nature Science Foundation of China(No. 61571236, 61320106008, 61602255), the Macau Science and Technology Fund(FDCT 093/2014/A2, 041/2017/A1), the Research Committee of University of Macau(MYRG2015-00011-FST, MYRG2015-00012-FST), Postgraduate Research & Practice Innovation Program of Jiangsu Province (KYCX17_0795).
Hao Gao is an Associate Professor with the Institute of Advanced Technology, Nanjing University of Posts and Telecommunications, Nanjing, China. Currently, He is a Visiting Fellow with the Department of Computer Science, University of Macau, Macau. His current research interests include Pattern Recognition and Artificial Intelligence with several publications in those fields.
References (25)
- et al.
Real-time geometry, albedo, and motion reconstruction using a single RGB-D camera
ACM Trans Graphics
(2017) - et al.
Brain Intelligence: go beyond artificial intelligence
(2017) - et al.
Multi-scale context for scene labeling via flexible segmentation graph
Pattern Recognit
(2016) - et al.
Fast reference frame selection based on content similarity for low complexity HEVC encoder
J Visual Commun Image Represent
(2016) - et al.
Gbest-guided artificial bee colony algorithm for numerical function optimization
Appl Math Comput
(2010) - et al.
An adaptive and hybrid artificial bee colony algorithm (aABC) for ANFIS training
Appl Soft Comput
(2016) A study on particle swarm optimization and artificial bee colony algorithms for multilevel thresholding
Appl Soft Comput
(2013)- et al.
Modified artificial bee colony based computationally efficient multilevel thresholding for satellite image segmentation using Kapur's, Otsu and Tsallis functions
Expert Syst Appl
(2015) - et al.
A quick artificial bee colony (qABC) algorithm and its performance on optimization problems
Appl Soft Comput
(2014) - et al.
A local search-based GeneSIS algorithm for the segmentation and classification of remote-sensing images.
IEEE J Sel Top Appl Earth Obs Remote Sens
(2016)
A new level-set-based protocol for accurate bone segmentation from CT imaging
IEEE Access
Multi-step radiographic image enhancement conforming to weld defect segmentation
IET Image Process
Cited by (0)
Hao Gao is an Associate Professor with the Institute of Advanced Technology, Nanjing University of Posts and Telecommunications, Nanjing, China. Currently, He is a Visiting Fellow with the Department of Computer Science, University of Macau, Macau. His current research interests include Pattern Recognition and Artificial Intelligence with several publications in those fields.
Zheng Fu is pursuing Master degree in Pattern recognition and intelligent system from Nanjing University of Posts and Telecommunications, Nanjing, China. Her current research interest include Pattern Recognition and Artificial Intelligence.
Chi-Man Pun is currently an Associate Professor and the Head of the Department of Computer and Information Science with the University of Macau. He has investigated several funded research projects and authored or coauthored over 100 refereed scientific papers. His research interests include digital image processing; multimedia forensics and watermarking; pattern recognition and computer vision.
Haidong Hu is a senior scientist at Beijing Institute of Control Engineering, China. He obtained his doctoral degree from HIT, China in 2009. He has finished his postdoctoral research at Tsinghua University, China in 2011. He is doing his visiting research at University of Missouri, United States. His areas of research interest are artificial intelligence and robotic vision.
Rushi Lan is a postdoctoral researcher at the School of Computer Science & Engineering, South China University of Technology, and also is an Assistant Professor at the School of Computer Science and Information Security, Guilin University of Electronic Technology. His research interest is computer vision and machine learning.
- ☆
Reviews processed and recommended for publication to the Editor-in-Chief by Associate Editor Dr. Huimin Lu.