Noise Induced Segmentation of Noisy Color Image Onkar Krishna

1, Rajib Kumar Jha

2, Anil Kumar Tiwari

1, Badal Soni

3

Indian Institute of Technology Jodhpur1, Ropar

2, NIT Silchar

3, India

Email: {onkarkris@gmail.com}

Abstract— In this paper we have proposed Noise Induced HSI

model based noisy and blurred colour image segmentation

technique. This approach uses additive noise to suppress the

effect of internal noise present in an image for proper detection of

objects from such images. In this algorithm we decompose a

given image in Hue, Saturation and Intensity (HSI) components

and then apply processing on intensity component of the

decomposed image. We measured performance of proposed

algorithm in terms of correlation coefficient and number of

mismatch pixels. The effectiveness of the proposed algorithm is

compared with the different existing techniques. It is observed

that the computational complexity of our algorithm is less in

comparison with several existing techniques, because it deals only

with intensity component of the decomposed image.

Furthermore, an additional advantage, our technique of

segmentation gives better performance as compared to SSR

based segmentation using RGB model, SR-extended, integrated

region matching, watershed and marker controlled watershed

based segmentation method.

Keywords-Image segmentation; superathreshold stochastic

resonance; noise; stochastic resonance

I. INTRODUCTION

Stochastic resonance (SR) is a noise induced phenomenon. SR occurs when noise enhances a distorted or noisy signal in a nonlinear system. The system’s performance such as signal-to-noise ratio, cross-correlation, or mutual information can increase when the moderate amount of noise is added to the system. This is so because that the system has nonzero-noise optimality. In order to exhibit SR, system should have following three properties: a nonlinearity, subthreshold signal and source of additive noise. SR occurs in physical and biological systems such as ring lasers [1], threshold hysteretic Schmitt triggers circuits [2], and A/D converters [3]. SR also occurs in biological systems such as rat [4].

There are number of techniques available for image segmentation but none of them are suitable for identifying target objects in noisy and blurred color images at different brightness levels. In the noise induced phenomenon, noise is added deliberately to the input noisy image to improve the performance of segmentation. This phenomenon highlights the utilization of noise and overcome the misconception about the destructiveness of the noise. A significant number of applications of this technique in image and signal processing have been reported. Jha et al. [6] proposed improved performance of watermark detection using suprathreshold stochastic resonance (SSR). In this work authors introduced SSR based method which improved the watermark detection performance in terms of correlation coefficient. In [7] author proposed an image denoising method using SSR. In this work, author applied white gaussian noise to obtain de-noised image for a given input noisy image. Similarly, same authors in [8]

proposed to use noise induced method for contrast enhancement of images. Peng et al. [9] proposed SR system based preprocessing approach for enhancement of medical images. Here system performance is improved by adding some suitable noise to the input signal. In [12] author proposes adaptive stochastic resonance for color object segmentation is reported in this paper

Color image segmentation is one of the most challenging task; it is important step in several application areas such as object classification, feature extraction and computer vision. Color thresholding is a simple yet effective algorithm that can segment objects in controlled lighting conditions. But in many applications lighting conditions of acquired images do not match with the preset thresholds. These deviated conditions can greatly worsen the performance of the color thresholding based segmentation algorithms. We can adapt thresholds for each image’s lighting conditions but the algorithm will become complex.

In this paper we proposed SSR based segmentation in which we generate N number of noisy frames from given noisy and blurred image. Apply Otsu’s method for bi-level thresholding on intensity values of HSI decomposed image, combine all frames using OR operation and we iterated this process for different standard deviation of input noise. We mention some existing color image segmentation algorithms separately in Section III and from experimental result we found that these algorithms are not suitable to get effective object boundary for noisy images also performance of these algorithms degrades for the objects having similar background.

The rest of paper is organized as follows Section II is about noise induced phenomenon. Section III reviews the object segmentation algorithms. Section IV is about proposed segmentation algorithm and it explains the performance enhancement when the right amount of noise is added to the system. Section V lists the performance measures that we use to show the SR effect. Section VI is about results and discussions followed by conclusions in Section VII and important references in Section VIII

II. NOISE INDUCED PHENOMENON

Noise induced phenomenon is a form of stochastic resonance (SR) that occurs in arrays of identical threshold devices subjected to independent and identically distributed additive noise. In such arrays, SR occurs regardless of whether the signal is subthreshold or not, hence the name suprathreshold SR. The block diagram of this phenomenon is shown in Fig.1 This phenomenon is quite general and is not restricted to any particular type of signal or noise distribution.

978-1-4673-5952-8/13/$31.00 ©2013 IEEE

1

1

( , ) ( , )N

i

i

R x y I x y=

= ∪

Figure 1: Block diagram of SSR phenomena

Fig.1 of n no of thresholded devices with their responses

( 1 nY Y− ), iY is response of th

i device that adds input signal X,

the noise iξ in the system is subject to additive noise ( iξ ).

The signaliY is unity if sum of the signal X and the8 noise

iξ

is greater than threshold ∆ . The output signal is zero otherwise.

The thresholded signal iY is summed to obtain the output

'Y

which is result of SSR phenomenon.

III. COLOR OBJECT SEGMENTATION ALGORITHM

A. Color Object Thresholding

This approach uses the RGB color space. A single plain-color object is segmented with a set of six threshold values, two for each of the three color dimension in the RGB color space. Thresholding is performed as shown below.

'1,

( , )0 ,

l ui f y

R x yo th erw ise

∆ ≤ ≤ ∆=

The algorithm repeats at all coordinate value (x,y) belonging to the given image and it is applied with all the three color components. Output of these process are binary images correspond to red, green and blue component of image. On applying AND operation on these binary images, the possible objects are obtained.

B. SR-extended Segmentation

Adaptive stochastic resonance for color object segmentation is reported by Janpaiboon et al. [12] in 2006. This method uses N stages of noisy RGB color thresholding. Each stage simply adds independent white Gaussian noise to a noisy input image before performing the color thresholding. Threshold values are selected manually using histogram plot of individual component of color image. Binary output images of all stages are combined with an OR operation to obtain a binary output image.

C. Region Extraction and Object Identification

This algorithm is used to identify a target object represented by a region of connected pixels in a binary image. From the RGB thresholded image, connected pixels region are found using connected pixels component analysis.

D. SSR Based Image Segmentation in RGB Color Model

In his method authors [10] generate N noisy RGB frames of zero mean and one standard deviation. Apply Chao bi-level thresholding to get N thresholded frames of each color component, then uses OR operation to get one thresholded frame of each color component, finally combined all three frames by taking AND operation. Repeat the above process at different standard deviations. The propose algorithm in [10] deals with all R, G and B frames simultaneously that worsen the performance and complexity becomes very high.

IV. PROPOSED NOISE INDUCED COLOR IMAGE

SEGMENTATION IN HSI MODEL

The eye is to be composed of rods and cones. Rods respond only to the intensity of the light falling on them, and are more sensitive to low light levels than cones.

What we perceive as color seems to depend on characteristics of brightness, hue, and saturation. We generally regard the basic colors as Red, Green, and Blue, and define other colors as a mixture of these three. There are several models used to describe the tristimulus color scheme: RGB, CMY (K), YIQ, and HSI. Each model was derived for specific purposes and has certain advantages over the others. Converting between the different models is generally done by a relatively simple mapping.

This section describes the proposed SSR based image segmentation. The input image is taken as multi-object, multi-colored image having different color backgrounds. In the proposed algorithm we have taken HSI color model. The algorithm steps are given below.

Step-1 N number of different noise frames with Gaussian noise of mean zero and some initial standard deviation are generated. The generated Gaussian noise frames are added to intensity component (I) of the HSI decomposed color input noisy image. Thus we have N different frames of I components added with noise frames. Now we apply thresholding on these N frames. Threshold value is obtained using Otsu’s bilevel thresholding approach. Noise addition for segmentation of input noisy image is tested with different number of noisy frames. This is because; as the number of noise frames increases the segmentation performance improves. However, we have to select the optimum number of noise frames to reduce computation time.

Step-2 now we have N number of thresholded frames. We logically ORed all the frames to obtain one frame. This operation provides maximum connected region of input noisy blurred color image correspond to initial standard deviation. The resultant image which is segmented binary output corresponds to noisy image. The mathematical representation of above is given below.

2 2

( ) ( ) ( )

x y

E xy E x E yC orrela tionC oeffic ient

σ σ

−=

Step-3 Step-1 and Step-2 is repeated with noise of different standard deviations giving n different frames

( 1 nR toR )

Step-4 Finally, ORing the entire n frames provides the final segmented output. The logical OR operation between n frames for different noise standard deviation provides the maximum common as well as the maximum connected regions. It has been observed experimentally that use of n=18 provide optimum results.

1

( , ) ( , )n

i

i

F x y R x y=

= ∪

Step-5 Connected component analysis is applied on resultant image to get final segmented output.

Step-6 Correlation coefficient and percentage of mismatched pixels are used for quantitative performance analysis of the proposed segmentation technique. High correlation coefficient and low number of extra pixels as well as change in object position provides correct segmented image. These parameters are calculated for full image.

V. PERFORMANCE MEASURE

A. Correlation Coefficient

Correlation coefficient is a measure of how well the output response is similar to the input. The general mathematical expression for correlation coefficient can be given as shown below.

Where x and y are actual input and output response signal. E(x)

and E(y) are expected values of x and y, 2

xσ

2

yσ

are

respective variances.

B. Number of Mismatched Pixels

This measure directly describes mistaken pixels between two binary images. The bit-wise XOR operation shows the pixels in which their binary values do not match to pixels in another binary image at the same locations. So we can obtain the amount of the error pixels by counting the results of 1 of the bit-wise XOR.

VI. SIMULATION AND RESULTS

Here we have tested the proposed SR segmentation algorithm

on four types of test images. Two images shown in Figure 4(a)

and (g) were generated by different noise and illuminations

from Test image-1 in Figure 2(a). In the same way we

generate Figure 5(a) and (g) from Test image-2 shown in

Figure 3(a). Figure 2(b) and Figure 3(b) are reference

segmented outputs, all performance measures are calculated

with respect to them. The noises in all stages were

independent Gaussian noise introduced through matlab10.0,

and blurring introduced using Adobe software.

We used four different segmentation algorithms to compare

with our proposed segmentation algorithm; these are Marker-

controlled watershed segmentation [11], SR-extended method

for segmentation [12], Watershed segmentation [5], and SSR-

based segmentation [10]

Our experiment was to find optimum number of frames (n

in Step-1 of Section IV). It is observed that at approximately

18 frames the performance measures become optimum.

Hence subsequent results are presented for n=18 and range of

noise variance is from 0.04 and 0.15. SSR based segmentation

algorithm using Gaussian noise in HSI model applied on test

images Test-1.1, Test-1.2, Test-2.1 and Test-2.2. Segmented

output is shown in Figure 4(b), Figure 4(h) and Figure 5(b),

Figure 5(h) by proposed technique. This result is compared

with the segmentation outputs using SSR method [13] (Figure

4(c) and Figure 4(i)), SR-extended (Figure 4(f) and Figure

4(l)), Watershed algorithm (Figure 4(d) and Figure 4(j)) and

Marker controlled watershed (Figure 4(e) and Figure 4(k)).

From all these visual outputs it is quite evident that noise

induced segmentation technique in HSI domain gives better

result compared to other techniques.

Figure 2(a) Test image-1 Figure 2(b) Reference segmented output

Figure 3(a) Test image-2 Figure 3(b) Reference segmented output

(a) Test-1.1 (Variance=0.04) (b) Segmentation by proposed algorithm (c) Segmentation by SSR (d) Watershed Segmentation

(e) MCW segmentation (f) Segmentation by SR-extended (g) Test-1.2 (Variance=0.07) (h) Segmentation by proposed algorithm

(i) Segmentation by SSR (j) Watershed Segmentation (k) MCW segmentation (l) Segmentation by SR-extended

Figure 4: Segmentation output of Test1.1 and Test1.2 image using our proposed algorithm, SSR, watershed, MCW, SR-extended methods

(a) Test-2.1 (Variance=0.04) (b) Segmentation by proposed algorithm (c) Segmentation by SSR (d) Watershed Segmentation

(e) MCW segmentation (f) Segmentation by SR-extended (g) Test-2.2 (Variance=0.07) (h) Segmentation by proposed Algorithm

(i) Segmentation by SSR (j) Watershed Segmentation (k) MCW Segmentation (l) Segmentation by SR-Extended

Figure 5: Segmentation output of Test2.1 and Test2.2 image using our proposed algorithm, SSR, watershed, MCW, SR-extended methods

TABLE I: Performance measure on test imagesTest1.1, Test1.2, Test2.1and Test2.2 using 18 no. of frames on an average and comparison with other

different methods

VII. CONCLUSIONS

Noise induced color image segmentation in HSI color model

for image segmentation is presented in this paper. Gaussian

noise has been used in this phenomenon. The proposed

segmentation technique has been applied on noisy images

under different intensity levels and blurring. The segmentation

outputs of our algorithm have been compared with outputs of

existing techniques such as SR-extended, Watershed and

Marker controlled watershed. Statistical performance

measures, such as, percentage of mismatched pixels and

correlation coefficient with respect to reference segmented

image has been used for comparison purpose. It has been

observed that percentage of mismatched pixels is much less at

the same time the correlation coefficient is much higher for

proposed method. Main advantage of our proposed algorithm

that, it perform better than existing algorithms for input noisy

color image and also our algorithm is efficient it deal with

only intensity component of a color model. The same

improvement in segmentation is observed for all images with

different levels of noise and under different intensity

conditions. Also visually, the outputs with our proposed

technique are better than other techniques. This establishes the

robustness of our technique against noise and intensity

variation.

VIII. REFERENCES

[1] B. McNamara, K. Wiesenfeld, and R. Roy, “Observation of

stochastic resonance in a ring laser,” Physics Letter A, vol. 60, no.

25, pp. 2626-2629, Jun. 1988.

[2] S. Fauve and F. Heslot, “Stochastic resonance in a bistable

system,” Physics Letters A, vol. 97, no. 1, 2, pp. 5-7, Aug. 1983.

[3] D. F. Russell, L. A. Willkens, and F. Moss, “Use of behavioral

stochastic resonance by paddle fish for feeding,” Nature, vol. 402, pp.

291-294, Nov. 1999.

[4] W. N. Lie, “Automatic Target Segmentation by Locally Adaptive

Image Thresholding”, IEEE Transaction on Image Processing, vol. 4,

pp. 1036-1046, Jul. 1995.

[5] J. B. T. M. Roerdink and A. Meijster, “The watershed transform:

definitions, algorithms, and parallelization strategies,” In Fundamenta

Informaticae, vol. 41, pp. 187-228, 2000.

[6] R.K. Jha, P. K Biswas and D. Mishra, “Improved watermark

detection performance using suprathreshold stochastic resonance,” In

TENCON2010, 2010.

[7] R.K. Jha, P. K Biswas and B.N Chatterji, “Image denoising using

stochastic resonance,” Proceeding of international conference on

cognition and recognition, 2011

[8] Rajib Kumar Jha, P. K. Biswas, B. N. Chatterji, “Contrast

Enhancement of Dark Images Using Stochastic Resonance,” Journal

of IET Image Processing, Vol. 6, Issue 3, Pages 230-237, April 2012.

[9] Renbin Peng, Hao Chen, Pramod K. Varshney and James H.

Michels, “Stochastic Resonance: An Approach for Enhanced Medical

Image Processing,” Life Science Systems and Applications

Workshop, 2007.

[10] R.K. Jha, P. K Biswas and B.N Chatterji, “Image segmentation

using suprathreshold stochastic resonance,” World academy of

science and technology, 2011.

[11] Kyung-Seok SEO, Chang-Joon PARK, Sang-Hyun CHO,

Heung-Moon CHOI, “Context-Free Marker-Controlled Watershed

Transform for Efficient Multi-Object Detection and Segmentation,”

Publication IEICE TRANSACTIONS on Fundamentals of

Electronics, Communications and Computer Sciences, vol. E84-A,

no.6, pp. 1392-1400, 2001.

[12] Sittichote Janpaiboon and Sanya Mitaim, “Adaptive Stochastic

Resonance in Color Object Segmentation,” IEEE International Joint

Conference on Neural Networks, vol. 1, pp. 2508-2515, 2006.

Methods

Test-1.1

Test-1.2

Test-2.1

Test-2.2

Correlation

Percentage

mismatched

Pixels

Correlation

Percentage

mismatched

Pixels

Correlation

Percentage

mismatched

Pixels

Correlation Percentage

mismatched

Pixels

Proposed

Algorithm

0.9641

1.53

0.9538

1.49

0.8957

3.20

0.8887

3.12

SSR

[10]

0.9065 3.45 0.8898 3.30 0.8224 8.65 0.8220 8.45

SR-extended

[12]

0.8974

3.03

0.8865

3.00

0.8025

9.89

0.7865

9.80

Watershed

[5]

0.8596

6.32

0.8475

6.28

0.7665

13.81

0.7543

12.60

MCW

[11]

0.8756

6.12

0.8656

6.03

0.7816

11.39

0.7714

11.29