Documente Academic
Documente Profesional
Documente Cultură
(1)
Where gray levels’ total count is represented by L. This generally ranges between (0-255), for a gray
scale image. The total pixels in the image is given by n, and is the associated probability for each of i
pixels. The Cumulative Distribution Function (CDF) corresponding to is given by the following equation:
(2)
Using the cumulative distribution function for transformation, HE maps the given input images’ intensity
values to values ranging between x0 to xL-1. Thus, the transformation function based on CDF is given by:
(3)
(4)
(5)
Therefore, the rate of enhancement can be limited by restricting the value of p(x). When the slope of the
histogram is clipped, the number of bins in the histogram is either increased or decreased based on the clip
parameter [29]. The histogram’s slope of the input image is trimmed using the clip limit parameter which is
given by equation(6)
(6)
where is the pre-defined clip limit and w is the size of tile or window region under consideration. For
a gray-scale image, the number of bins is 256, i.e., the range of pixels from 0-255. Those regions that are
clipped are then redistributed to the existing bins. HE is then computed for the altered histogram [30]. Though
CLAHE provides an edge over AHE by providing a limit factor for minimizing amplification of noise, the
downside of CLAHE method is that the histograms are trimmed at a pre-fixed user specified clipping value.
This value remains the same for all images and is crisp. Hence, the proposed research brings in the idea of
applying Triangular Fuzzy Membership (TFM) Function for determining clipping limit automatically.
3.4. Proposed Methodology -Triangular Fuzzy Membership Contrast Limited Adaptive Histogram
Equalization (TFM) CLAHE algorithm
TFM-CLAHE provides automatic, image variant fuzzy clip-limit for limiting contrast and resulting in
enhanced image. The proposed research applies Triangular Fuzzy Membership (TFM) Function for determining
clipping limit and performs clipping accordingly. The clipped portion of histogram that surpasses the clip limit
are redistributed among all histogram bins equally. The triangular fuzzy membership function (TFM) takes three
values as inputs which form the minimum, maximum and mean values of image’s pixel intensities in the
considered window and computes the fuzzy number which determines the clipping parameter for image
I. The value returned by TFM ranges between 0 and 1 and clipping is done accordingly. Those intensity values
that exceed the calculated clipping limit are redistributed as in CLAHE, thus resulting in a smoothened
histogram.
(7)
Where the parameters {a, b, c} (with a < b < c) determine the coordinates of the three corners
of the underlying TFM. Here denotes the image dependant clipping parameter.
The point b, with membership value of 1, is the mean value and a and c are left hand spreadand right
hand spread of .
The TFM-CLAHE algorithm computes the fuzzy clip-limit by using the following equation, thus is a
replacement of pre-defined crisp clip-limit with a TFM calculated fuzzy clip limit.
(8)
This TFM lets areas of lower local contrast to attain a higher contrast and limits noise in high contrast
regions. After computation of clip-limit, the calculation of probability of occurrence of pixel is given
as follows:
(9)
(10)
This becomes the image’s accumulated histogram. The transformation function is given by
(11)
(12)
……
(iv)Compute clip
(v) Limit the contrast based on
limit using (TFM)
TFM clipping parameter
(vi)Check
Required
Contrast
No
Yes
(i) (ii)
(iii) (iv)
(v)
Figure 2: Comparison of Iris Image and Histograms (i) Original unenhanced image (ii) Image Enhanced
using HE (iii)Image Enhanced using AHE (iv) Image Enhanced using CLAHE (v) Image Enhanced using
Proposed TFM-CLAHE
The results of experiments are evaluated on three traits of biometrics, namely Iris, Face and Fingerprints
from these databases. Multiple modalities are considered for evaluating the effectiveness of the proposed
algorithm on heterogenous images and compared against classical enhancement techniques. Average
Information Content (AIC), Naturalness Image Quality Evaluator (NIQE), Peak Signal Noise Ratio (PSNR) and
Mean Square Error (MSE) are used as metrics to assess the effectiveness of the proposed algorithm.
Figure 2 shows gray scale images of Iris and their corresponding histograms. Results of enhanced images
using HE, AHE, CLAHE and proposed TFM-CLAHE are also sketched. For the histogram images, intensity
values from 0-255 are plotted in x-axis since images under consideration are gray scale images. The count of
each pixel values is plotted in y-axis. Figure 2(i) shows the original iris image and its respective histogram. In
2(ii) iris image is enhanced using histogram equalization. Though the original image is enhanced, the
enhancement is uniform. In case of an iris recognition system, the region of interest is at the centre and higher
contrast in these regions will yield better results rather than uniform enhancement. Histogram of HE enhanced
image shows uniform distribution of intensity values. Figure 2(iii) shows the results of enhancing given input
image using Adaptive Histogram Equalization (AHE) technique. AHE, though comes as a solution to global
enhancement, this also results in over amplification of unwanted noise in local regions. The corresponding
histogram shows pixel intensities that are distributed non-uniformly. Iris image enhanced using CLAHE is
sketched in figure 2(iv). Though amplification of noise in local regions are limited, using a pre-defined clip limit
yields a uniform enhancement. Histogram of CLAHE enhanced image also shows peak values for certain
intensity values. Image enhancement using TFM-CLAHE is shown in figure 2(v). The figure shows that the
proposed method exhibits better enhancement over other methods. The figure also illustrates that the region of
interest shows higher contrast values than other regions. The circular shaped pupil, iris and surrounding circular
regions have higher brightness values than other regions of the image. An iris image enhanced in such a
way,tend to produce better results for further processing. The histogram of the proposed method also exhibits
smoother distribution of pixel intensities.
In figure 3, the results of enhancement of various techniques applied on facial images are illustrated. The
original image 3(i) is enhanced using Histogram Equalization (HE) and the outcome is shown in figure 3(ii).
The facial features are enhanced in a global way, affecting pixel intensities to be distributed in a smoother way.
Such an enhancement leads to a washed-out effect of the image. AHE enhancement as shown in figure 3(iii)
contributes to a good amount of contrast enhancement. The natural appearance of image is compromised in the
process of enhancement .The histogram of these images exhibits peak values between high and low pixel
intensity levels. Limiting the contrast using TFM computed clip limit shows better enhancement in terms of
preserving the brightness. Figure 3(v) shows TFM-CLAHE enhanced image which appears visually pleasing
and natural as compared to figure 3(iv) which is the resultant of enhancement using CLAHE where the contrast
is limited based on a pre-defined limit. The histogram of the proposed method also shows that the pixels are
evenly spread and distribution is smoother.
Figure 4 shows the comparative results of images enhanced using classical techniques and proposed
technique. The techniques are applied on fingerprint images. Ridges, valley, delta, cross-over regions present in
finger prints make these traits unique and differ from individual to individual. HE enhanced fingerprint applies a
global enhancement and the regions like ridges, delta etc. are uniformly enhanced along with other regions of
fingerprint. This is depicted in figure 4(ii). The histogram corresponding to this image shows smooth
distribution. Figure 4(iii) shows image of fingerprint enhanced using AHE. Contextual regions are fairly
enhanced, but unnecessary artefacts are also enhanced. The histogram of AHE enhanced image also shows few
peak values for certain pixel values. CLAHE enhanced image shown in figure 4(iv) exhibits good enhancement
though clear demarcation of ridges and valleys is subtle. The corresponding histogram also shows non-smooth
distribution. Image enhancement using proposed TFM-CLAHE is shown in figure 4(v). Fingerprint image is
enhanced by limiting the contrast at a clip-limit that is computed according to the given fingerprint image.
Enhanced TFM-CLAHE image shows clear demarcation among valleys, ridges, delta and cross-overs. The pixel
distribution in the corresponding histogram is also smooth and uniform.
Thus, the results reveal that the proposed method, TFM-CLAHE, provides better enhancement on varied
images that are used. The clip-limit varies according to the image and limits the contrast appropriately. The
region of interest in each of these images are suitably enhanced. The main aim of the proposed research, TFM-
CLAHE is to enhance the contrast of images by applying enhancements to local regions and keeping the
contrast level below image-dependant clip-limit. The results also show that the proposed method retains the
naturalness of images and are visually pleasing. Adding on to this, the region of interests shows higher contrast
values than other regions. The cause of such appropriate enhancement is use of right clip-limit that is not pre-
determined but is adjusted based on the image. Thus, the use of triangular membership function to compute
clipping limit for varied images, yield better results since there is no pre-determination of any clip-limit and can
be varied automatically depending on the image under consideration.
(i) (ii)
(iii) (iv)
(v)
Figure 3: Comparison of Lena Face Image and Histograms (i) Original unenhanced image (ii) Image
Enhanced using HE (iii) Image Enhanced using AHE (iv) Image Enhanced using CLAHE (v) Image
Enhanced using Proposed TFM-CLAHE
(i) (ii)
(iii) (iv)
(v)
Figure4: Comparison of Fingerprint Image and Histograms (i) Original unenhanced image (ii) Image
Enhanced using HE (iii) Image Enhanced using AHE (iv) Image Enhanced using CLAHE (v) Image
Enhanced using Proposed TFM-CLAHE
4.1 Image Quality Assessment (IQA)
Average Information Content (AIC), Natural Image Quality Evaluator (NIQE), Peak Signal to Noise
Ratio (PSNR) and Mean Square Error (MSE) are the metrics that are employed in the proposed research work to
assess quality of enhanced image. These values are also compared with classical image enhancement
algorithms. IQA provides mathematical models that evaluate the quality of images.
4.1.1 Average Information Content (AIC)
AIC computes the information content of an image. Image with a high AIC value, shows that the
information content of the image is also high. Hence an image that yields a higher value is said to carry higher
information content. The average information content or entropy is defined as
(13)
Figures 5, 6 and 7 illustrate AIC values for images enhanced using proposed TFM-CLAHE technique
and image enhancement techniques such as HE, AHEand CLAHE. The proposed technique is applied on
datasets from three different databases. It is apparent from the figures that the images enhanced using the
proposed method carry high information content as compared to other techniques.
Figure 5: Comparison of AIC values for CASIA Iris Database among various image enhancement
techniques and proposed algorithm
Figure 6: Comparison of AIC values for CASIA Face Database among various image enhancement
techniques and proposed algorithm
Figure 7: Comparison of AIC values for CASIA Fingerprint Database among various image enhancement
techniques and proposed algorithm
AIC
Biometric IMG
Trait # Original TFM-
HE AHE CLAHE
Image CLAHE
1 5.6949 4.8625 6.8025 5.2026 7.1877
2 5.7597 4.91 6.7215 5.314 7.0041
Iris 3 5.7631 4.8854 7.0232 5.4341 7.3042
4 6.702 5.6077 7.0564 6.2693 7.5077
5 6.355 5.2081 7.4393 5.8552 7.5277
6 7.2525 5.9714 7.0116 7.1805 7.4872
7 7.3234 5.6938 7.2536 7.1584 7.4009
Lena Face 8 6.5275 5.3296 7.1853 6.3067 7.2593
9 6.8398 5.7256 7.022 6.612 7.4888
10 7.1898 5.9594 7.2705 7.0098 7.4792
11 7.2769 5.8814 7.0602 6.888 7.6455
12 7.1583 5.8787 7.1777 6.6138 7.638
Finger Print 13 6.9607 5.8625 7.2572 6.5259 7.6533
14 6.735 5.6383 7.1303 6.3325 7.3788
15 6.604 5.6722 7.2695 6.2326 7.4258
Table 2: Comparison of AIC values
Table 2 shows AIC values of original image and images enhanced using HE, AHE and CLAHE. The
values are computed for three different set of images namely iris, face and fingerprint. It is evident from the
table that the proposed TFM-CLAHE method has higher AIC values when compared to other techniques.
4.1.2Natural Image Quality Evaluator (NIQE)
Yet another metric for measuring image quality is Natural Image Quality Evaluator (NIQE) [31-32]. It is
based on a model of Natural Scene Statistic (NSS). This is based on simple space domain which includes
collection of statistical features. The roots of these images are from a set of images that are both natural and
undistorted. The NIQE index scores range between 0 and 100 with lower values indicating better image quality.
Figures 8, 9 and 10 show NIQE values computed on proposed method and conventional techniques. The
outcome of calculation shows that TFM-CLAHE enhanced images have lower NIQE values in comparison with
other techniques.
Figure 8: Comparison of NIQE values for CASIA Iris Database among various image enhancement
techniques and proposed algorithm
Figure 9: Comparison of NIQE values for CASIA Face Database among various image enhancement
techniques and proposed algorithm
Table 3 summarizes the natural image quality evaluator index of images enhanced using various
techniques and the values show that the image enhanced using the proposed TFM-CLAHE method exhibits
better image quality.
Figure 10: Comparison of NIQE values for CASIA Fingerprint Database among various image
enhancement techniques and proposed algorithm
NIQE
Biometric IMG
Trait # Original TFM-
HE AHE CLAHE
Image CLAHE
1 21.8744 20.887 22.8775 18.8784 17.8709
2 19.8759 20.8784 18.8809 18.8797 17.2165
Iris 3 22.8765 21.8785 21.8788 19.8347 18.7807
4 19.8777 19.8791 18.8801 18.4127 17.8807
5 21.8752 22.8777 21.8774 22.8785 19.8785
6 4.2582 4.6226 4.92 3.0971 3.0262
7 4.5149 5.2806 4.7013 4.3492 3.7423
Lena Face 8 4.9694 3.4264 4.8383 4.163 3.098
9 6.4476 4.2088 5.298 5.4924 4.0194
10 6.8143 6.6968 5.5055 6.4829 5.4829
11 22.8869 19.884 22.887 21.8934 18.7644
12 18.8798 18.8873 20.8891 19.8871 18.7092
Finger Print
13 21.8804 20.8779 20.8816 20.8933 20.4302
14 19.8831 19.8809 19.8785 18.8919 17.3209
15 21.8894 20.8869 21.0028 20.8927 19.8875
Table 3: Comparison of NIQE values
4.1.3 Mean Square Error (MSE)and Peak Signal to Noise Ratio (PSNR)
Mean Square Error is an image quality assessment metric. It assesses the error sensitivity of the enhanced
image as compared to the original image. Lower values of MSE indicate lower error in the enhanced image as
compared to the original. It is given by the following equation.
(14)
where is the gray level values of original image O and is the gray level values of enhanced
image E. M and N are the dimensions of the images.
Peak Signal to Noise Ratio (PSNR) is a significant metric that measures the quality achievement of
enhanced image in comparison with the original image as well as it indicates the degree to which contrast in the
image is achieved. A higher value of PSNR indicates that there is a minimal degradation of the image as
compared to the original image. PSNR computation is given by the following equation.
(15)
where L is the intensity level of image and for a gray scale image, it is 255.
Figures 11 to 13 illustrate the comparative values of MSE. The values of computed MSE for images
enhanced using various enhancement techniques and proposed method are tabulated in table 4. The table and
figures indicate that the images enhanced using the proposed method exhibits lower error values as compared to
the original image, thus yielding a better quality.
Figure 11: Comparison of MSE values for CASIA Iris Database among various image enhancement
techniques and proposed algorithm
Figure 12: Comparison of MSE values for CASIA Face Database among various image enhancement
techniques and proposed algorithm
Figure 13: Comparison of MSE values for CASIA Fingerprint Database among various image
enhancement techniques and proposed algorithm
Figures 14 to 16 illustrate PSNR values of images enhanced using various image enhancement
techniques. These values are tabulated in table 5. The table and figures indicate that the images enhanced using
the proposed method exhibit higher values of PSNR as compared to the original image, thus revealing that the
images are enhanced with better contrast and quality.
MSE
Biometric IMG
Trait # HE AHE CLAHE TFM-CLAHE
1 0.00052 0.00170 0.00014 0.00006
Figure 15: Comparison of PSNR values for CASIA Face Database among various image enhancement
techniques and proposed algorithm
Figure 16: Comparison of PSNR values for CASIA Fingerprint Database among various image
enhancement techniques and proposed algorithm
PSNR
Biometric IMG
Trait # HE AHE CLAHE TFM-CLAHE
1 32.8397 27.6627 38.54 42.2291