Unimodal thresholding
Introduction
Over the years, many image thresholding techniques have been developed [1], [2], and considerable research continues at present [3], [4], [5]. The reason for this long term, ongoing effort is that none of the methods are capable of optimal performance under all conditions. Although techniques based on many different principles are available, they all operate under certain implicit or explicit assumptions. These assumptions effectively form restrictions on the successful operation of the thresholding algorithms. Examples of requirements for bilevel thresholding are that there should be two distinct modes in the intensity histogram, the peaks should not be too dissimilar in size, and they should be roughly Normal. Since different computer vision applications produce different types of images we can expect the success of any thresholding algorithm to vary according to how well its expected operating conditions are met.
This paper looks at the case where bilevel thresholding is desired even though the image histogram may only contain one obvious peak. Although performing thresholding might not seem appropriate in such cases there are in fact many instances when it is required. Take for example edge detection, which typically produces many low-magnitude edges corresponding to non-edges. The true edges produce a wide range of edge magnitudes that will often just create a fat tail on the non-edge peak in the edge histogram rather than generate a distinct peak of their own. Another example is change detection carried out by differencing pairs of images. In applications such as surveillance the amount of change will be extremely small if a wide field of view is imaged and only a small, distant object has moved. Thus the histogram could contain one enormous peak (no-change) and one tiny peak (change).
Such applications will cause difficulties for most existing thresholding algorithms, and there has been relatively little work in such areas [6], [7]. In contrast, we present in this paper a technique that is specifically intended to cope with essentially unimodal distributions rather than the more usual bimodal or multimodal distributions. That is, the second peak is either very small, or is submerged within the main peak. Section 3 describes the algorithm and analyses its performance on a variety of types of synthetic data. In Section 4 it is applied to thresholding a range of types of real data: edges, corners, difference images, optic flow, texture difference images, polygonal approximation of curves, and image segmentation. Both the theoretical and experimental analyses demonstrate the benefits of the new algorithm for such data.
Section snippets
Previous work
There is a wide range of bimodal and multimodal image thresholding techniques available [1], [2]. Early methods often analysed the shape of the histogram, looking for concavities as suitable threshold points [8]. Others were based on the statistics of the classes. For instance, for bilevel thresholding Otsu [9] minimises the ratio of the between class variance to the within class variance of the two classes. Relaxation labelling has also been applied to thresholding: pixels are assigned initial
Proposed method
The proposed bilevel thresholding algorithm is extremely simple. It assumes that there is one dominant population in the image that produces one main peak located at the lower end of the histogram relative to the secondary population. This latter class may or may not produce a discernible peak, but needs to be reasonably well separated from the large peak to avoid being swamped by it. A straight line is drawn from the peak to the high end of the histogram. More precisely, the line starts at the
Examples
Having demonstrated the potential strengths of the proposed algorithm on synthetic data we now show its application to several thresholding tasks, this time involving real data. Histograms are generated from a variety of sources. To reduce the effects of noise and quantisation that are most evident on small data sets the number of histogram bins in each application is set to , where P is the number of image pixels and G is the number of grey levels.
Discussion
A simple thresholding algorithm has been described. Unlike the majority of thresholding algorithms it is suitable primarily for essentially unimodal distributions. Tests on synthetic data showed the feasibility of the approach. As long as (1) the mode is not so broad as to fill most of the histogram, and (2) the mode is not too strongly peaked (e.g. it is not exponential) then the mode is likely to contain a corner at its base that can be detected and is also a suitable threshold. Further tests
Acknowledgements
I would like to thank Geert van Kempen and Lucas van Vliet for bringing Zack et al.'s paper to my attention.
About the Author—PAUL ROSIN received the B.Sc. degree in Computer Science and Microprocessor Systems in 1984 from Strathclyde University, Glasgow, and the Ph.D. degree in Information Engineering from City University, London in 1988. He was a research fellow at City University, developing a prototype system for the Home Office to detect and classify intruders in image sequences. He worked on the Alvey project “Model-Based Interpretation of Radiological Images” at Guy's Hospital, London before
References (29)
- et al.
A survey of thresholding techniques
Comp. Vision, Graphics Image Process.
(1988) - et al.
Fast recursive algorithms for 2-dimensional thresholding
Pattern Recognition
(1998) - et al.
Minimum cross-entropy threshold selection
Pattern Recognition
(1996) - et al.
A new method for grey-level picture thresholding using the entropy of the histogram
Comp. Vision, Graphics Image Process.
(1985) A fast thresholding selection procedure for multimodal and unimodal histograms
Pattern Recognition Lett.
(1995)An iterative procedure for the polygonal approximation of plane curves
Comp., Graphics Image Process.
(1972)Moment-preserving thresholding
Comp. Vision, Graphics Image Process.
(1985)- et al.
Texture classification using texture spectrum
Pattern Recognition
(1990) An analysis of histogram-based thresholding algorithms
Graphic. Models Image Process.
(1993)- et al.
Adaptive thresholding by variational method
IEEE Trans. Image Process.
(1998)
Image thresholding using fuzzy entropies
IEEE Trans. Systems Man Cybern. B
Edges: saliency measures and automatic thresholding
Mach. Vision Appl.
The analysis of cell images
Ann. NY Acad. Sci.
Cited by (0)
About the Author—PAUL ROSIN received the B.Sc. degree in Computer Science and Microprocessor Systems in 1984 from Strathclyde University, Glasgow, and the Ph.D. degree in Information Engineering from City University, London in 1988. He was a research fellow at City University, developing a prototype system for the Home Office to detect and classify intruders in image sequences. He worked on the Alvey project “Model-Based Interpretation of Radiological Images” at Guy's Hospital, London before becoming a lecturer at Curtin University of Technology, Perth, Australia, and later a research scientist at the Institute for Remote Sensing Applications, Joint Research Centre, Ispra, followed by a return to the UK, becoming lecturer at the Department of Information Systems and Computing, Brunel University London, UK. Currently he is senior lecturer at the Department of Computer Science, Cardiff University.
His research interests include the representation, segmentation, and grouping of curves, knowledge-based vision systems, early image representations, machine vision approaches to remote sensing, and the analysis of shape in art and architecture. He is on the executive committee of the British Machine Vision Association.