Image Segmentation of Wounds and Their Performance Evaluation (Image Segmentation)

NONLINEAR FILTERS AS EDGE ENHANCERS AND THEIR APPLICATION TO WOUNDS

Ms. A.G. Deshpande, Dr. T. R. Sontakke, Ms. M.S. Joshi

Assistant Professor (Electronics), Govt. Engg. College, Aurangabad 431005., Professor (Electronics), SGGSCE Tech, Nanded 431602. Assistant Professor (Computer), Govt. Engg. College, Aurangabad 431005

INTRODUCTION Abstract

This paper presents computerized analysis of wounds(injuries) in which different edge detection techniques are used like Sobel, Prewit, Robert , Laplacian. Their performance evaluation is done. Parameters for performance evaluation are Edge Detection Error Ratio (EDER). To improve the performance of edge detectors , nonlinear edge enhancers are used as pre filters for edge detectors. Those filters are able to convert smooth edges to step edges and suppress noise simultaneously, so false alarms due to noise are minimized and edge gradient estimates tend to be larger and localized which improves edge maps and performance of edge detector.

Keywords: Wounds (Injuries), edge detection , EDER, two dimensional distance.

Today, computers are introduced in all branches of Biomedical Engineering. Looking into utilization of computers, a virgin field is presented here, ‘FORANSIC MEDICINE AND TOXICOLOGY’. In general the computers can be used in many ways but record maintenance and analysis of Medico-Legal cases is major work, where computer come into picture.

Major concept of analysis of Medico-Legal cases is, computer-assisted analysis of wounds (injuries).

Wounds [1], [2] are defined as “ Disruption of normal structure and tissues by application of force to body”. Types of wounds are

1)Abrasion,

2) Contusion,

3) Laceration,

4) Incised wound,

5) Stab wound,

6) Fire Arm,

7) Fracture.

Nature of wounds [2] depends on

1. Nature of weapon
2. Amount of energy transferred.
3. Condition under which energy is discharged.
4. Nature of tissues affected.

Police Department has to deal with analysis of wounds and weapons, for this they usually take help from doctors i.e. medico-legal experts from Forensic Medicine and Toxicology Department. Doctors do their work manually, like deciding the shape of wound, based on that weapon used, is detected. As a human being there are certain limitation and person-to-person variations as well.

More over wounds are not permanent structures. It either gets cured or decays. Many times the status of wound is stored in form of theoretical description or inform of photographs. In Medico - Legal cases, the details of wounds are presented in court of law, where the cases prolong for many years. Under such condition permanent storage is preferred.

An attempt has been made to develop a Computer Vision System, which includes a computer, a video grabber (or a camera and scanner). Of course a computer cannot replace a medico-legal expert but can assist definitely. So total idea can be explained as photographs of actual wounds are taken, scanned and stored into computer in digitized form, which becomes a permanent record and can be used for further analysis.

First step of analysis is segmentation. Segmentation includes thresholding, ANN based segmentation , Fuzzy based segmentation. Numbers of thresholding techniques are implemented and their performance evaluation is carried out [3]

Another way is, Boundary Detection, which requires edge detection [ 5]. Experimental results with animals and human beings shows that edges are one of the most important clues for interpreting images. Edges constitute a significant portion of information contents in image and thus it is useful to extract these features from a complete image.. An edge map is vastly reduced complexity and retains important structure present in original image. The most common edge detection algorithms are those based on the gradient , such as Sobel , Prewit , Robert , and Laplace operator.

Performance of edge detector can be compared in many ways. Canny [ 6 ] has proposed three parameters. Basically in noiseless case , all operators are roughly equivalent but do not perform well on images that are noisy. Quantitatively , performance of edge detectors can be calculated by EDER [5], Two Dimensional Distance [9] and Chi – Square Test [7].

To improve the performance of edge detectors , it is useful to pre filter the image prior to performing edge detection [ 4 ]. Smoothing the filter are commonly used as pre filters when images are corrupted by high levels of noise. For example, a median filtering can be used to remove impulse . However these filters tend to reduce edge gradients. While linear sharpeners can increase edge gradient, but generally they are unable to produce step edges and tend to be extremely sensitive to noise

Here it is proposed that , nonlinear edge enhancing filters can be used as prefilters for gradient based edge detectors. It is shown that the performance of gradient based edge detectors can be significantly improved by utilizing the non linear sharpner. These filters have common properties that make them effective in this application , like they are capable of converting blurred edges to ideal step edges and can suppress noise at the same time, these filters do not introduce ringing at edges. Thus false alarms are minimized. And gradients tend to be large and localized. This leads to thinner edge contours and reduced sensitivity to threshold value. Also edges that would go undetected are easily detected after prefiltering.

Section II describes non linear filter. In section III comparative measures are defined. Section IV gives results of edge detection before and after prefiltering and their effect on performance evaluation parameters. Finally section V describes conclusions.

II NONLINEAR SHARPNERS

The nonlinear filters considered here are lower-upper-middle (LUM) filters [ 4 ]. Before presenting paper some notations are prescribed.

Consider ,a d-dimensional discrete sequence { x (n) } where the index n=[ n1,n2,n3…..,nd]. Also, consider a moving window that spans N samples at each location, where N is assumed to be odd. These samples can be indexed and written as a vector x(n)= [ x1(n),x2(n)……, xN(n) ]. The middle sample in the observation window is denoted x(N+1)/2(n) and the filter estimate at this location is denoted y(n).

The rank ordered or sorted observation samples are written as

x1(n) <=x2(n)<=…… xN(n).

1. The LUM Filter

The output of the LUM filter with parameters k and l is given by

| x (k) , if x (N+1) / 2 < x (k)

| x (l) , if x (l) < x (N+1)/2 <= tl

Y=FLUM (x)=| x (N-l+1) , if tl < x (N+1) / 2 < x (N-l+1)

| x (N-k+1) , if x (N-k+1) < x (N+1)/2

| x (N+)/2 , otherwise

Where tl is the midpoint between x (l) and x ( N-l+1l) and l <= k <= l<=(N+1)/2. The parameters k and l can be considered tuning parameters that allow the LUM filter to have a variety of characterstics. In case where l=(N+1)/2 and k is varied , LUM filter acts as a smoothing filter . if k=l = (N+1)/2 then LUM acts as sharpner . As l is decreased , x ( N-l+1l) and x ( l) move towards the upper and lower extreme values respectively. This leads to an increased edge enhancing effect. When 1< <k <= (N+1)/2, sharpening and outlier rejection can be achieved simultaneously. The parameter k can bee increased to improve the impulse rejection characterstics of the LUM filter . The parameter l , on the other hand , controls the level of edge enhancement. This parameter is decreased to give more enhancement . Finally , the filter performs an identity operation when k=l and l=(N+1)/2.

III PERFORMANCE MEASURES

Canny [ 6 ] has proposed few performance of edge detector, like SNR, Localization Error etc but these can be implemented on standard image signals only like step edge etc . For natural images, those parameters cannot be implemented.

Edge detectors can be compared in a number of ways . Basically, eye itself performs some sort of edge detection . But still some quantitative evaluation is essential. Performance of edge detector is measured by parameters like Edge Detector Error Rate(EDER), Two Dimensional Distance [9] and Chi Square Test [7].

EDER can be calculated in form of edge detector error rate, which is defined as

EDER=N(1)/N(0)

Where N(0) be the number of edge pixels declared. And N(1) be the be the number of missing or added edge pixels after adding noise. And N(0) is held fixed for the noiseless as well as noisy image .

Two Dimensional Distance is defined as the two dimensional distance between expected output and actual output [9]. In this case, it is calculated between Edge detected image before adding noise and after adding noise. It is evaluated as given below

x y

Two Dim Distance = Sqrt (( A(i,j)-M(i,j))2

i=1 j=1

Where, Ai = Edge detected image before adding noise

Mi = Edge detected image after adding noise to image and prefiltering it.

x*y = size of image.

Chi-square Test [7] , which gives discrepancy existing between actual output and expected output. Like Two Dimensional Distance , even this parameter is calculated between Edge detected image before adding noise and after adding noise. It is evaluated as given below

n

2=(Oi-Ei)2/Ei

i=1

Oi, Ei= Populations of ith gray level of Edge detected image before adding noise and after adding noise and pre filtering respectively.

All these parameters indicate the discrepancy between actual and ideal values so it is very much desirable that all parameters should have low values.

IV EXPERIMENTS AND RESULTS

In this section we present outcome of several edge detection experiments. MatLab is used as programming platform for implementation. These programs can be applied on any type of wound images. For implementation we have taken an image of wound produced by bullet (Fire Arm).

In first stage of analysis, all edge detectors i.e. Sobel , Prewit, Robert, Laplacian are operated on wound images. (Figure (1.1) is selected for analysis)

Then images are corrupted with Gaussian noise with variance=0.01. EDER is calculated. Table 1 indicates performance of all edge detector on the basis of EDER , Two Dimensional Distance and Chi Square Test .

Table 1

Figure (1.1)

Figure (1.2)

Figure (1.3)

Figure (1.4)

Figure( 1.5 )

Figure (1.6)

First stage of analysis includes calculation of EDER for operators like Sobel, Prewit, and Robert, Laplace operator. Fig (1.1) represents original image,

Fig (1.2) represents corrupted image,

Fig (1.3) represents Sobel edge detected image,

Fig (1.4) represents Prewit edge detected image,

Fig (1.5) represents Robert edge detected image,

Fig (1.6) represents Laplacian edge detected image,

For demonstrating the effect of nonlinear filtering on EDER, we have selected commonly used Sobel edge detector.

The original image has been corrupted with a Gaussian point spread function and has been processed using nonlinear filter i. e. LUM. A 3x3 window with k=l=3 and N =9 is used for LUM filter.

at threshold=130.

Fig (2.1) represents original image

Fig (2.2) represents corrupted image,

Fig (2.3) represents Sobel edge detected image,

Fig (2.4) represents Sobel edge detected image without pre filtering,

Fig (2.5) represents Sobel edge detected image with LUM pre filtering

Figure 3 represents EDER with and without pre filtering. Graph with LUM indicates edge map with pre filtering. It is lower as compare to graph without pre filtering for all threshold values. At the same time EDER for LUM is zero at threshold 130 whereas EDER for without pre filtering is nowhere zero, which indicates the sharpening effect of LUM.

V CONCLUSION

In the noiseless case all operators are roughly equivalent but in presence of noise their behavior is different. A close look at Table 1 indicates that among all the edge operators Sobel and Prewitt are better as compare to Robert and Laplacian. Regarding Robert, even though it is giving lowest values of all parameters, the visual appearance is not good. About Laplacian neither EDER nor visual appearance are up to the mark. Graph of EDER Vs Threshold indicates that prefiltering improves the EDER of Sobel operator significantly. It is a good combination of noise suppression and edge enhancement properties make the nonlinear filter very much suitable for the application . Visually also Sobel map with pre filtering and Sobel map without pre filtering indicates the effect of it.

Figure (3)

REFERENCES

[1] M. K. Krishna Reddy ‘ Hand book of Forensic Medicine’ pp 90, M/s Medical Publishers, Delhi India

[2] Dr. C. K Parikh,’ Parikhs Text book of Jurisprudence and Toxicology for Classrooms and Courtrooms’, pp 236-297, M/s Medical Publishers, Delhi India

[3] Prof. Smt. A. G. Deshpande, Dr. T. R. Sontakke ‘Objective Evaluation of Image Thresholding techniques for Wounds’, 12 the International Conference on Pattern Recognition, 27-28 June 2002 Portugal,

[4] Gradient based edge detection using nonlinear edge enhancing prefilters – Russel C Hardie, Charles G. Boncelet. , IEEE on IP 4/ 11/ nova 95

[5] A.K Jain, “ Fundamentals of Digital Image processing , PHI, Third Edition , 1997.

[6] John canny, ‘A Computational Approach to Edge Detection’, IEEE Tran on PAMI, Vol – PAMI – 8, no 6, 1986

[7] Murray Spiegel “Book theory and problems of statistics” 2nd edition

[8] W. K. Pratt , Digital Image Processing, Wiley , New York 1978

[9] Pattern recognition and Image analysis – Earl Gose , Richard John, Jost , PHI, 1999