A NEW ITERATIVE TRICLASS THRESHOLDING TECHNIQUE IN IMAGE SEGMENTATION

ABSTRACT

Image processing, segmentation is often the first step to pre-process images to extract objects of interest for further analysis. Segmentation techniques can be generally categorized into two frameworks, edge-based and region-based approaches. As a segmentation technique, Otsu’s method is widely used in pattern recognition, document binarization, and computer vision. Otsu’s method is used as a pre-processing technique to segment an image for further processing such as feature analysis and quantification. Otsu’s method searches for a threshold that minimizes the intra-class variances of the segmented image and can achieve good results when the histogram of the original image has two distinct peaks, one belongs to the background, and the other belongs to the foreground or the signal. The Otsu’s threshold is found by searching across the whole range of the pixel values of the image until the intra-class variances reach their minimum. In this paper, we present a new iterative method that is based on Otsu’s method but differs from the standard application of the method in an important way. At the first iteration, we apply Otsu’s method on an image to obtain the Otsu’s threshold and the means of two classes separated by the threshold as the standard application does.

EXISTING SYSTEM:

Otsu’s method, the new method starts to create better results as we can compare. And at the final iteration, the new method creates much better results as shown., indicating that in the process of searching for optimal TBD regions for segmentation, the new method is not adversely affected by some low distance ratios at initial iterations. We showed the histogram of the original and the changes of the thresholds. The thresholds monotonically decrease from a large value of 121.0 on iteration one, which is also the Otsu’s threshold, to a small value of 52.8 on iteration seven, indicating that the iterative method progressively searches for weak objects to segment.

PROPOSED SYSTEM:

The original image and its Otsu’s result are shown respectively. We observe that the standard Otsu’s method misses the weak nucleus at the bottom of the image. We then applied the iterative method and show the result of the first and fourth iterations, which is the last iteration. Experiments on another microscopic image with multiple nuclei. (a) A zoomed-in portion of muscular nuclei acquired by fluorescence microscopy, the weak nuclei are marked by red arrow. (b) The result of Otsu’s method. Note it missed the weak nuclied pointed by the arrows in (a).(c) Results after the first iteration of the new method. (d) Results after the fourth iteration of the new method. The nuclei missed by the standard Otsu’s method are now detected by the proposed method. We note that although the final result does not fully segment the whole areas of the weak nuclei, starting with the result we can recover the nuclei by using techniques such as region-regrowing in post-processing. respectively. While the first iteration also misses the weak nucleus, the algorithm detects the nucleus at the final iteration. As the last example, another microscopic image consisting of many nuclei with various gray intensities was tested. The original image contains some weakly stained nuclei, as pointed by the arrows. Result given by Otsu’s method shows that the method misses weak nuclei. By applying the proposed method on the same image, we found that the weak nuclei are gradually detected by the iterative process. At the final result, all three weak nuclei are marked correctly. Here we note that though the final segmentation result given by the new method do not reveal the full sizes of the three weak nuclei, they can be readily recovered in post processing by techniques like region growing. In both examples, the iteration stops when the change in threshold T [i] is less than two. We also computed the distance ratios and show the plots respectively. From the figure we observe that despite very low distance ratios at the first iterations for both figures, the ratios quickly increase from iteration two, suggesting that the TBD regions at each iteration provide favorable inputs for Otsu’s method to process. The above examples on both synthetic and real images demonstrate that the new method achieves superior performance in segmenting single or multiple objects, even in very challenging cases.

IMPLEMENTATION

Implementation is the stage of the project when the theoretical design is turned out into a working system. Thus it can be considered to be the most critical stage in achieving a successful new system and in giving the user, confidence that the new system will work and be effective.

The implementation stage involves careful planning, investigation of the existing system and it’s constraints on implementation, designing of methods to achieve changeover and evaluation of changeover methods.

MODULES

A. Otsu’s threshold Method

B. Iterative Method

MODULES DESCRIPTION :

A. Otsu’s threshold Method

Otsu’s method searches the histogram of an image to find a threshold that binarizes the image into two classes, the background with a mean of μ0 and the foreground with a mean of μ1, as shown in the top .Without loss of generality, here we assume that the foreground is brighter than the background. The calculation of threshold T is as follows. we can see that T is function of the pixel values of both the foreground and the background. If the signal intensity changes, it may affect T in such a way that the segmentation result may become less optimal. Next we use an example to illustrate the effect of signal intensity on the calculation of T. Fig. 2(a) shows an original image consisting of multiple objects in gray scale. The segmentation result of the standard Otsu’s method is shown in Fig. 2(b), from which we can observe that most objects are correctly segmented or marked. Then we purposely added a strong object to the original image to increase the overall signal intensity in the foreground and tested how Otsu’s method performs in this case. The new test image is shown in Fig. 2(c) and the corresponding Otsu’s result is shown in Fig. 2(d). Though one would expect that with added signal intensity the segmentation result should be equally good or better, the result of Fig. 2(d) shows that some weak objects are actually missed now by Otsu’s method. As shown above, there are cases that Otsu’s method does not produce satisfactory results even when the foreground has a high signal intensity, i.e., a higher signal-to-background ratio (SBR). In other words, the performance of Otsu’s method is not a function of SBR only.

B. Iterative Method

The idea of dividing an image’s histogram iteratively into three classes is illustrated at the bottom of Fig. 1. For an image u, at the first iteration, Otsu’s method is applied to find a threshold T [1] where the superscript denotes the number of iteration. We then find and denote the means of the two classes separated by T [1] as μ[1] 0 and μ[1] 1 for the background and foreground, respectively. Then we classify regions whose pixel values are greater than μ[1] 1 as foreground F[1] and regions whose pixel values are less than μ[1]0 as background B[1]. For the remaining pixels u(x, y) such that μ[1]0≤ u(x, y) ≤ μ[1]1 we denote them as the TBD class _[1]. So our iterative process assumes that the pixels that are greater than the mean of the “tentatively” determined foreground are the true foreground. Similarly, pixels with values less than μ0 are for certain the background. But the pixels in the TBD class, which are the ones that typically cause misclassifications in the standard Otsu’s method, are not decided at once and will be further processed.

CONSLUSION

As Otsu’s method is widely used as a pre-processing step to segment images for further processing, it is important to achieve a high accuracy. However, since Otsu’s threshold is biased towards the class with a large variance, it tends to miss weak objects or fine details in images. For example in biomedical images, nuclei and axons may be imaged with very different intensities due to uneven staining or imperfect lightening conditions, raising difficulty for algorithms like Otsu’s method to successfully segment them. Without a robust segmentation results, more sophisticated processing such as tracking and feature analysis become highly challenging. In this paper, we proposed to take advantage of Otsu’s threshold by classifying images into three tentative classes instead of two permanent classes in an iterative manner. The three classes are designated as the true foreground and background, and a third TBD region that is to be further processed at the next iteration. At each iteration, the tri-class approach keeps regions that are determined to be foreground and background unchanged and focuses on the third TBD region. At each succeeding iteration, the area of the TBD region decreases and more pixels are assigned to the foreground and background classes. The iteration stops until the change in thresholds of two consecutive iterations is less than a threshold. To assist on evaluating the performance of the new algorithm we introduced the notion of distance ratio which measures a posteriori how favorable an image or region is for Otsu’s method to segment. The performance of the new algorithm is evaluated on both synthetic and real microscopic images. By assigning very strong and very weak pixels to the tentative foreground and background classes, the new method is less biased toward the class with a large variance than Otsu’s method does. Experimental results demonstrate that the proposed algorithm can achieve superior performance in segmenting weak objects and fine details. The new method is also almost parameter-free except for the preset threshold to terminate the iterative process. The added computational cost is minimal as the process usually stops in a few iterations and each iteration only processes a monotonically shrinking TBD region. From a statistical analysis perspective, Otsu’s method is optimal to separate a bi-modal histogram into two classes where the probability distribution functions (PDFs) of the two classes have an approximately “tall and thin” shape. However, when one or both PDFs have a “wide and flat” shape a single threshold determined by Otsu’s or other method may not be sufficient to correctly separate the two classes as some pixels in the signal class may appear closer to the noise class on the histogram and become difficult to segment. Indeed, in Otsu’s method, as well as in many segmentation methods, a common challenge is to decide which pixels should or should not be binarized by a common threshold. If we have this information, then we can design perfect segmentation algorithm for virtually no errors. Testing results show that the new method can achieve better performance in challenging cases. We note that there are many segmentation methods, but many of them require careful selection of parameters to achieve satisfactory performance. From this perspective, a parameter-free method may be well suited in many applications.

SYSTEM SPECIFICATION

Hardware Requirements:

•  System : Pentium IV 2.4 GHz.

•  Hard Disk : 40 GB.

•  Floppy Drive : 1.44 Mb.

•  Monitor : 14’ Colour Monitor.

•  Mouse : Optical Mouse.

•  Ram : 512 Mb.

Software Requirements:

•  Operating system : Windows-7 32 bit Ultimate OS.

•  Coding Language : C#.Net

•  Front-End : Visual Studio 2010 Ultimate.

•  Data Base : SQL Server 2008.