Computers and Electronics in Agriculture 151 (2018) 196–206

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Original papers

Image dehazing based on dark channel prior and brightness enhancement T

for agricultural remote sensing images from consumer-grade cameras
Jiawei Zhanga,b, Xiuyuan Wanga,b, Chenghai Yangc, Jian Zhangc,d, Dongjian Hea,b,
⁎
Huaibo Songa,b,c,
a
Northwest A&F University, College of Mechanical and Electronic Engineering, 22 Xinong Road, Yangling, Shaanxi 712100, China
b
China Key Laboratory of Agricultural Internet of Things, Ministry of Agriculture, Yangling, Shaanxi 712100, China
c
USDA-ARS Southern Plains Agricultural Research Center, 3103 F and B Road, College Station, TX 77845, United States
d
Huazhong Agricultural University, College of Resource and Environment, 1 Shizishan Street, Wuhan, Hubei 430070, China

A R T I C LE I N FO A B S T R A C T

Keywords: Remote sensing technology has been widely used for monitoring crop fields and other agricultural applications.
Dehazing However, the clarity of remote sensing images is often affected by clouds and chaotic media in the atmosphere.
Remote sensing image Image dehazing can be achieved through the dark channel prior method (DCP), but there is always a brightness
Image enhancement distortion problem after image dehazing. To address the problem, this study proposed an improved image de-
Dark channel prior
hazing approach based on the DCP method and determined optimal enhancement parameters. Four evaluation
indices, including mean square error (MSE), peak signal to noise ratio (PSNR), average gradient and program
running time, were first calculated to evaluate the quality of enhanced images. An example image was dehazed
by the DCP method initially using the four indices to determine optimal dehazing parameters. Results showed
that image enhancement achieved the best effect when the dark channel window size Ω(x) is 5, atmospheric light
A is 215/255, and the lower limit t0 of transmission factor t(x) is 0.1. Next, these indices were applied to evaluate
the enhancement methods used in this research. The logarithmic enhancement method was finally selected as
the optimal method with the base number (1 + r) = 11 and enhancement parameter m = 0.5. To verify the
effectiveness of the selected method, 50 airborne images from a consumer-grade camera flown by an agricultural
aircraft were used to evaluate the improved method. Both the original and the enhanced images after dehazing
were mosaicked by Adobe Photoshop software. The mosaicked images before and after image dehazing were
compared. Results showed that the mosaicked image without dehazing had an entropy of 6.359 and an average
gradient of 6.513. In comparison, the mosaicked image with dehazing had an entropy of 6.668 and an average
gradient of 11.305, which were 4.86% and 73.58% higher than the respective values for the mosaicked image
without dehazing. These results indicate that the proposed method in this study is effective and can be applied to
dehaze remote sensing images.

1. Introduction compromised to some extent (Ansia and Aswathy, 2015). As atmo-

spheric scattering is related to distance from the target to cameras, the
Remote sensing technology has been widely used in agriculture, degree of image degradation changes with space (Iqbal et al., 2016).
forestry, geology, environmental protection as well as many other Therefore, in order to reduce influences of these factors on images and
areas. However, remote sensing images are always affected by adverse to improve the reliability of remote sensing data, image dehazing
weather conditions. The clouds in the atmosphere are one of the main technology has become a practical and valuable research topic and has
influencing factors (Riaz et al., 2016). As visible remote sensing images attracted increasing attention (Ni et al., 2016).
are mainly affected by clouds and atmospheric turbid media such as Fog image enhancement methods can be divided into two cate-
molecules and water droplets, the images are always degraded (Shi gories: globalized and localized. Globalized fog image enhancement
et al., 2014). In addition, the images captured even in sunny weather refers to adjusting the gray values determined by statistical information
conditions will also be affected because of the atmospheric scattering. of the whole image, regardless of the area where the adjusted point is
Image contrast and color fidelity of the degraded images will also be located (Luan et al., 2017; Xiao and Gan, 2012; Ge et al., 2015). For

⁎
Corresponding author at: Northwest A&F University, College of Mechanical and Electronic Engineering, 22 Xinong Road, Yangling, Shaanxi 712100, China.
E-mail address: [email protected] (H. Song).

https://doi.org/10.1016/j.compag.2018.06.010
Received 25 October 2017; Received in revised form 1 June 2018; Accepted 4 June 2018
0168-1699/ © 2018 Elsevier B.V. All rights reserved.
J. Zhang et al. Computers and Electronics in Agriculture 151 (2018) 196–206

Fig. 1. Research site and flight route map.

example, image brightness preserving ability was improved by the 2. Materials and methods
modification of traditional histogram equalization (Dhal and Das,
2017). Homomorphic filtration was used as a preliminary data pro- 2.1. Materials
cessing method for upgrading the flaw images obtained during coherent
echo-signal processing (Badalyan and Bazulin, 2003). A wavelet-based 2.1.1. Study sites
algorithm for image enhancement was proposed to solve the problem of This study was conducted at a cropping area of approximately
noise over-enhancement efficiently (Wu et al., 2004). Qureshi and 38.9 km2 near College Station, Texas, on July 14, 2015. The geo-
Deriche (2016) proposed a wavelet algorithm for image compression graphical location for the center of the area is 30°31′34.97″ N and
based on compressive sensing. A low-light image restoration method 96°24′30.44″W. The study site is adjacent to the Brazos River near
based on the variational Retinex model was put forward by Park et al. College Station, Texas and has an abundant vegetation distribution,
(2017). However, the degree of degradation in fog scenes is related to making it suitable for this study. Due to extensive cloudy and humid
its depth, and images often contain complex depth information. As a weather in May to July, aerial imagery acquired during this period is
result, globalized image enhancement is often not the ideal method. always affected by haze, making it difficult for crop monitoring.
Localized fog image enhancement operates on a portion of the entire
image. The transformation or transfer function is determined based on 2.1.2. Airborne image acquisition
the statistics of the region. For example, the Kuwahara edge corner A low-cost, dual-camera imaging system consisting of two Nikon
keeping filter was used to estimate atmospheric scatter light and the D90 digital cameras was used in this study. One camera equipped with
filter was further improved to achieve a more accurate dielectric AF Nikkor 24 mm f/2.8D lens was used to capture three-band RGB
transmittance (Wang et al., 2013). He et al. (2011) put forth the dark images and the other was modified to capture near-infrared (NIR)
channel prior (DCP) method and used image statistical methods to images. To obtain consistent images, the cameras were set to the
obtain general law of fog images. Then the results were combined with manual mode and the lens focus set to infinity. In order to obtain high
known fog imaging models and interpolation methods to obtain high- quality images, exposure time, aperture opening, and ISO speed were
quality dehazing images. Verdenet et al. (1981) applied local area set to 1/1000 s, f/6.3 and 200, respectively. All other parameters were
histogram equalization to images with differences in intensity. Choi and set to defaults. The captured image had an array of 4288 × 2848 pixels.
Yun (2011) presented a contrast enhancement method using an ad- The imaging system was mounted on the right step of an Air Tractor AT-
vanced image formation model in which an image was divided into 402B. A wireless remote control was attached to the GPS receiver to
three components. automatically trigger the camera for image acquisition. To achieve an
Dehazing based on DCP theory could increase the depth of field overlap of at least 50% along and between flight lines, images were
effect to make the image more realistic. However, the brightness of the acquired at 10-s intervals with a flight speed of 240 km/h along eleven
dehazed image is always distorted (Guo et al., 2017). The objectives of flight lines spaced at 1066-m intervals. Images were stored in 12-bit
this research were to dehaze remote sensing images using an improved RAW (NEF format) and 8-bit JPEG files in a SD memory card. A set of
DCP method, to determine the optimal dehazing parameters through 50 continuously captured RGB images belonging to the first six adjacent
objective evaluation indices, and to compare the dehazed images with flight lines was selected for this study. The study site and the imaging
original images using selected evaluation indices. system are shown in Fig. 1. The yellow1 line is the flight route and the
red arrows are the starting and ending directions. Detailed descriptions
can be found in Yang and Hoffmann (2015).

1
For interpretation of color in Fig. 1, the reader is referred to the web version of this
article.

197
J. Zhang et al. Computers and Electronics in Agriculture 151 (2018) 196–206

Fig. 2. Flowchart of the algorithms used in this research.

2.1.3. Software and hardware where Jc represents each channel of the color image and Ω(x) refers to a
Image mosaicking was carried out by using Adobe Photoshop CC window centered on pixel X.
(Adobe Systems Incorporated, San Jose, California). Other image pro- The DCP theory points out that J’s dark primary colors Jdark should
cessing was performed using Matlab 2015a (MathWorks, Natick, tend to 0. According to the above definition, Eq. (4) is obtained:
Massachusetts). All of the algorithms in this study were implemented on
J dark (x ) = min[ min (J c (y ))] = 0
a laptop computer with the Intel i5 processor, 4G memory equipped c y ∈ Ω(x ) (4)
with Windows 7. c
A is always positive. Therefore, Eq. (5) is obtained:
2.2. Methods min [min(J c (y )/Ac )] = 0
y ∈ Ω(x ) c (5)

The flowchart of image processing is shown in Fig. 2, including Assuming that the transmittance t(x) is constant in each window
∼
parameter selection, initial dehazing, image enhancement, image and defined as t (x ) . And the value A is given. Then the minimum value
stitching, and stitching method selection. Among them, the optimal is calculated twice for Eq. (2) to obtain Eq. (6):
dehazing parameters such as window size, lower limit of pixel trans- ∼ ∼
min (min I c (y )/ Ac ) = t (x ) min (min J c (y )/ Ac ) + 1− t (x )
mittance threshold t0 and the best value of atmospheric light A were y ∈ Ω(x ) c y ∈ Ω(x ) c (6)
selected by the parameter selection module. To carry out the algorithm
Substituting Eq. (5) into Eq. (6), Eq. (7) is obtained:
mentioned below, the original images were dehazed initially using the
∼
DCP method. Then three image enhancement methods were used and t (x ) = 1−ω min (min I c (y )/ Ac )
the results were compared. By applying image evaluation indices, the y ∈ Ω(x ) c (7)
∼
optimal enhancement method was selected before image stitching. Fi- where t (x ) is the estimated value of transmittance and ω is the in-
nally, the optimal stitching method was selected using the same eva- formation with fog, whose assignment affects the final image of the
luation indices, and the stitched images after dehazing were compared dehazing effect. It can make the image look more real, with the depth of
with the original images with haze. field effect and it generally takes the value of 0.95.
If the value of the pixel transmittance t(x) of a pixel in the trans-
2.2.1. Introduction to dark channel prior algorithm mission map is small, it will cause the value of J to be too large, re-
According to the theory of machine vision (Xie et al., 2017) for sulting that the whole image is excessively white. Therefore, a threshold
computer vision and computer graphics usages, the fog pattern for- value t0 is set. When t(x) is less than t0, let t = t0. According to Eq. (1), if
mation model can be described by Eq. (1): the threshold is set to t0, the final recovery Eq. (8) could be obtained:
I (x ) = J (x ) t (x ) + A (1−t (x )) (1) I (x )−A
J (x ) = +A
max(t (x ), t0) (8)
where x refers to a pixel block of the image, I refers to intensity of the
observed image, J refers to intensity of the scene light, A refers to global where a typical value of t0 is 0.1 (Ding and Tong, 2013).
atmospheric light, and t(x) is the pixel transmittance. The goal of image Based on the theory above, a fog image obtained from Section 2.1
dehazing is to derive J from A, t, and I. Since it cannot be obtained was processed initially, as shown in Fig. 3. The left image is the example
directly by Eq. (1), He et al. (2011) proposed an image dehazing image affected by haze, and it can be seen that the image is blurred and
method based on DCP. some details are difficult to distinguish. The right image is the initially
Eq. (1) can be converted to Eq. (2): dehazed image using the DCP method. It can be seen that the dehazed
I c (x )/ Ac = t (x ) J c (x )/ Ac + 1−t (x ) (2) image becomes clearer but has obvious brightness distortion, which is a
problem to be addressed in this study.
where A represents each channel of the global atmospheric light, Ic
c

refers to the intensity of the image observed for each channel, and Jc 2.2.2. Brightness enhancement for remote sensing images
refers to the light intensity of each channel. Brightness enhancement can be divided into linear and non-linear
For any input image J, its dark channel Jdark is expressed by Eq. (3): methods (Russo, 2004). Linear enhancement generally converts the
J dark (x ) = min ( min J c (y )) RGB image into Hue, Saturation, Luminance or Hue, Saturation, Value
y ∈ Ω(x ) c ∈ {r , g , b} (3) (HSL/HSV) color spaces. After adjusting the HSL/HSV components, the

198
J. Zhang et al. Computers and Electronics in Agriculture 151 (2018) 196–206

(a) Example image with haze (b) Initial dehazing result (brightness distortion)
Fig. 3. Dehazing result using dark channel prior.

image is re-transferred into RGB space. The advantage of this method is communication theory. It is the ratio of the maximum signal to the
that the adjusted image hierarchy is strong. The disadvantage is that the noise intensity. Since digital image represents image pixels with dis-
code is complicated and the adjusting speed is slow. When the image crete numbers, the maximum pixel value of an image is used to replace
brightness increases or decreases significantly, there will be a large the maximum signal volume. The larger the maximum pixel value is,
image distortion (Tong and Chen, 2017). Non-linear enhancement is to the smaller the distortion is. PSNR is calculated according to Eq. (13).
add or subtract a value to RGB image pixels. The advantage is that the
PSNR = 20·log10 (255/ MSE ) (13)
code is simple and the adjusting speed is fast. The disadvantage is that
image information is lost, and the adjusted image appears dull and non- The average gradient is a directional change in the intensity or color
layered (Khan et al., 2016). in an image. The magnitude of this rate of change can be used to re-
Linear enhancement is to adjust all pixels with the same enhance- present the relative clarity of image. The larger the value is, the more
ment conditions, as is shown in Eq. (9). details the image has. The formula is shown in Eq. (14).
J (i, j ) = I (i, j ) + c (9) G (x , y ) = d x i + d y j
where J(i, j) is the enhanced image, I(i, j) is the image to be enhanced, c d x (i, j ) = I (i + 1, j )−I (i, j )
is the enhanced coefficients, and (i, j) is the pixel coordinate. d y (i, j ) = I (i, j + 1)−I (i, j ) (14)
Logarithmic transformation is a non-linear method. This transform
maps a narrowband low gray scale input image value into a wide band where I represents the image pixel value (such as the pixel value of
output value, as is shown in Eq. (10). RGB), and (i, j) are the pixel coordinates.
Information entropy is defined as the average amount of informa-
J = log1 + r (I + r ) (10) tion produced by a probabilistic stochastic source of data. The formula
is shown in Eq. (15).
where J is the logarithmic transformed image, I is the original image,
and (1 + r) is the logarithmic base. The greater the base is, the stronger 255

the emphasis on the low gray portion is. The compression of the high H= ∑ pij log pij
i=0 (15)
gray part is also stronger.
Exponential transformation is also a nonlinear transformation. The where i refers to grayscale value and j represents the mean of horizontal
formula is shown in Eq. (11). gray scale. The range of i and j is [0, 255]. pij represents the probability
J = Im (11) that the pixel block of gray scale i is adjacent to the pixel block with
gray scale j. The formula is shown in Eq. (16):
where J is the exponential transformed image, I is the original image,
and m is the exponential parameter. The gray scale can be compressed pij = f (i, j )/ N 2 (16)
by selecting the value of m reasonably (Navarro et al., 2013).
where f(i, j) is the frequency of the characteristic binary (i, j). The two-
dimensional information entropy of the image reflects the average
2.2.3. Objective evaluation indices
amount of information in the image, whose value range is [0, 8]. The
The mean square error (MSE), peak signal to noise ratio (PSNR),
larger the value is, the more uniform the gray value of the image would
entropy, and average gradient (Chen et al., 2016) are commonly used as
be.
image evaluation indices to measure the dehazing effects. Program
operation time is also included in the evaluation method.
The mean square error between the original distorted image and the 3. Experimental results and analysis
enhanced image is calculated to measure the degree of distortion of the
image. The larger the value is, the larger the degree of the distortion 3.1. Parameter selection
would be. MSE is calculated by Eq. (12).
3.1.1. Best value of Ω(x)
Nx Ny
Specific procedures for obtaining an image dark channel included
MSE = ∑∑ (f (i, j )−f ̂ (i, j ))2 / Nx Ny
window size input, analyzing all pixels of the image row by row under
i=1 j=1 (12)
the selected window size, and merging the dark channel based on the
where f(i, j) is the value of each pixel of the corresponding reference sequence of the original image (Wang et al., 2015). Therefore, the
image, f ̂ (i, j ) is the pixel value of the image after processing, Nx and Ny window size Ω(x) affects the dark channel image, which in turn affects
represent the number of rows and columns of the image matrix, re- the final dehazing effect.
spectively. In order to find the optimal window size, atmospheric light A took
PSNR is an indicator showing image quality. It is proposed based on 215/255 and the lower limit of the transmission factor t(x) was not set.

199
J. Zhang et al. Computers and Electronics in Agriculture 151 (2018) 196–206

(a) original image (b) (x)=1 (c) (x)=3 (d) (x)=5

(e) (x)=7 (f) (x)=13 (g) (x)=17 (h) (x)=21

Fig. 4. Dark channel images obtained under different windows sizes.

(a) (x)=1 (b) (x)=3 (c) (x)=5 (d) (x)=7

(e) (x)=13 (f) (x)=17 (g) (x)=21

Fig. 5. Final dehazing results.

Table 1 The image shown in Fig. 4(a) was selected as an example. The dark
Evaluation results of window size. channel images under different dark channel window processing
Window size Ω(x) MSE PSNR (dB) Average gradient Running time (s)
methods are shown in Fig. 4. It can be seen that as the window size
increases, the dark channel image becomes blurred and blocky.
1 1107 40.7345 9.4416 125 Fig. 5(a–g) show the final dehazing results of the corresponding
3 738 44.7865 9.4571 147 images in Fig. 4 by using the DCP method. It can be seen that the de-
5 588 47.059 9.671 183
7 495 48.7715 9.803 232
hazed images under these seven windows have better dehazing effect
13 347 52.3381 9.813 435 compared with original images. In Fig. 5, when the window size be-
17 306 53.573 9.8839 663 comes larger, image clarity tends to be better in general. The four ob-
21 290 54.1424 9.8294 973 jective evaluation indices previously introduced were used to evaluate
the dehazing images with the seven windows. It is noteworthy that the

Fig. 6. Window size Ω(x) evaluation results.

200
J. Zhang et al. Computers and Electronics in Agriculture 151 (2018) 196–206

(a) t0 = 0.1 (b) t0 = 0.2 (c) t0 = 0.3

Fig. 7. Dehazed images of different transmission factors.

Table 2 Table 3
Evaluation results of t0. Evaluation results of atmospheric light A.
Value of t0 MSE PSNR Average gradient Atmospheric light A MSE PSNR Average gradient

0.1 411 50.641 9.597 195/255 165 59.7364 9.3245

0.2 411 50.641 9.597 205/255 202 57.7178 9.454
0.3 411 50.641 9.597 215/255 411 50.6406 9.5964
225/255 817 43.7638 9.7376
235/255 1450 38.0342 9.8674
245/255 2339 33.2509 9.9888
fallow field in the middle of the image shows some patterns because of
ringing effect, but this will not affect the results.
Table 1 presents the evaluation results. It can be seen from Table 1
3.1.2. Best value of lower limit t0 of transmission factor t(x)
that as the dark channel window size increases, MSE, PSNR, and the
The evaluation experiment of the lower limit of the transmission
average gradient improve except that the average gradient is slightly
factor t(x) was carried out under the condition that the dark channel
worse when Ω(x) = 21, compared with Ω(x) = 17. Program running
window Ω(x) = 5 and the atmospheric light A = 215/255. Fig. 7(a–c)
time increases as the dark channel window increases. Especially after
show the output images for t0 = 0.1, 0.2, and 0.3 respectively. There
Ω(x) = 13, the running time increases significantly.
are no visible differences among these three images.
Fig. 6 shows different dehazing effects under different window size
Table 2 shows the value of MSE, PSNR and average gradient of
Ω(x) more intuitively. As program running time is very long and not
dehazing images with the three values of t0. The program running time
practical when Ω(x) = 21, it is not listed in Fig. 6.
was not included in this experiment because the values of t0 did not
From Table 1 and Fig. 6, it can be concluded that:
affect program running time. The experimental results showed that
lower limit of t0 had no effect on the dehazed images.
(1) No matter how large window size is, the dehazed images are better
Although the experimental images did not change with the value of
than the original images;
t0, the direct attenuation item J(x)t(x) will tend to 0 if t(x) tends to 0.
(2) MSE decreases with the increase of Ω(x);
The original image obtained by direct recovery tends to contain noise.
(3) PSNR and the average gradient increase slowly with the increase of
Therefore, it is necessary to set a lower limit t0 for the transmission
Ω(x);
factor t(x). In this research, a typical value of t0 = 0.1 was used.
(4) Program running time increases significantly after Ω(x) = 5.

Based on the above four factors, the best effect can be obtained 3.1.3. Best value of atmospheric light A
when dark channel window Ω(x) = 5. Fig. 8 shows the images obtained under the above conditions
(Ω(x) = 5 and t0 = 0.1) for different values of atmospheric light A:

(a) A = 195/255 (b) A = 205/255 (c) A = 215/255

(d) A = 225/255 (e) A = 235/255 (f) A = 245/255

Fig. 8. Image dehazing results under different values of A.

201
J. Zhang et al. Computers and Electronics in Agriculture 151 (2018) 196–206

Fig. 9. Evaluation results of atmospheric light A.

(a) c = 10 (b) c = 20 (c) c = 30

(d) c = 40 (e) c = 50 (f) c = 60

Fig. 10. Enhancement results with different values of c.

(a) (1+r) = 9 (b) (1+r) = 10 (c) (1+r) = 11

(d) (1+r) = 12 (e) (1+r) = 13 (f) (1+r) = 14

Fig. 11. Enhancement results with different (1 + r) values.

195/255, 205/255, 215/255, 225/255, 235/255 and 245/255. When The evaluation results of atmospheric light A are shown in Table 3
A = 195/255, there is visible haze that affects the image quality. and Fig. 9. Similarly, the program running time was not included in this
Therefore, values smaller than A = 195/255 were not considered. experiment by the same reason.
It can be seen from Fig. 8 that there are significant differences As illustrated in Table 3 and Fig. 9, as A becomes larger, MSE and
among the dehazed images with different values of A. As A becomes PSNR get worse. Especially when A = 215/255, MSE increases greatly.
larger, the dehazing effect is increasingly obvious. However, the overall PSNR also increases greatly with the increase of A, especially after
contrast of images becomes higher and the differences are subtle to A = 205/255. The average gradient improves with the increase of A,
distinguish. but there is little change.

202
J. Zhang et al. Computers and Electronics in Agriculture 151 (2018) 196–206

(a) m = 0.3 (b) m = 0.4 (c) m = 0.5

(d) m = 0.6 (e) m = 0.7 (f) m = 0.8

Fig. 12. Enhancement results with different values of m.

Table 4 It can be concluded from Table 3 and Fig. 9:

Linear enhancement evaluation results with different c values.
Enhancement parameter c Entropy Average gradient
(1) No matter how large A is, the dehazed images are better than the
original ones;
10 5.6174 3.9217 (2) MSE and PSNR get worse as A becomes large;
20 5.6174 3.9214 (3) The average gradient essentially does not change with A.
30 5.6174 3.9208
40 5.6173 3.9201
50 5.6172 3.9198 Based on the above three factors, it can be concluded that the best
60 5.6171 3.9186 dehazing effect can be achieved when A = 215/255.

3.2. Experiment results

Table 5
Logarithmic enhancement evaluation results with different (1 + r) values.
After obtaining the best dark channel window where Ω(x) = 5, at-
Base number (1 + r) Entropy Average gradient mospheric light A = 215/255 and transmission factor t(x) = 0.1, 50
airborne images were processed using the improved dehazing method.
9 6.3435 6.1583
10 6.3479 6.1586
The dehazed images were mosaicked using the Photomerge function in
11 6.3552 6.1620 Photoshop software.
12 6.3507 6.1449
13 6.3517 6.1392
3.2.1. Experiment results of linear enhancement
14 6.3496 6.1243
Since the enhancement parameter c in Eq. (9) is a variable, it is
necessary to find the best value of parameter c. Fig. 10 shows the en-
Table 6 hancement results obtained using the linear enhancement method. It
Exponential enhancement evaluation results with different m values. can be seen from Fig. 10 that as c becomes larger, the images are
brighter and the haze is heavier.
Exponent parameter m Entropy Average gradient

0.3 5.8069 4.1268 3.2.2. Experiment results of logarithmic transformation enhancement

0.4 5.9555 4.5208 Since the base number (1 + r) in Eq. (10) is a variable, the loga-
0.5 5.9931 4.6166
rithmic transformation of the image varies with (1 + r). Images en-
0.6 5.9659 4.4861
0.7 5.9043 4.2913 hanced by the logarithmic transformation method with different base
0.8 5.8172 4.2913 numbers are shown in Fig. 11. The best base number (1 + r) ranges
from 9 to 14 according to preliminary experiments. Therefore, only the
images with these base numbers are shown. It can be seen from Fig. 11
Table 7 that with the increase of the base number, the brightness value has
Evaluation results based on two nonlinear enhancement methods. increased slowly.
enhancement method Entropy Average gradient
3.2.3. Experiment results of exponential enhancement
Logarithmic enhancement with (1 + r) = 11 6.3552 6.1620
Images enhanced by the exponential enhancement method with
Exponential enhancement with m = 0.5 5.9931 4.6166
different exponential parameter m values are shown in Fig. 12. The best
values of exponential parameter m ranges between 0.3 and 0.8 ac-
cording to preliminary experiments. Therefore, only the images with
these values are shown.
It can be seen from Fig. 12 that as the exponent m increases, the
image brightness decreases. According to this observation, it can be

203
J. Zhang et al. Computers and Electronics in Agriculture 151 (2018) 196–206

(a) Automatic stitching (b) Perspective stitching (c) Cylindrical stitching (d) Spherical stitching
Fig. 13. Mosaicked images before dehazing using four methods.

(a) Automatic stitching (b) Perspective stitching (c) Cylindrical stitching (d) Spherical stitching
Fig. 14. Mosaicked images after dehazing using four methods.

found that the logarithmic transformation enhancement method has the is the best base number (1 + r) for the logarithmic transformation en-
potential to get optimal results. hancement method.

3.3. Result analysis 3.3.3. Objective evaluation of exponential transformation

Evaluation results for the exponential enhancement method with
3.3.1. Objective evaluation of linear enhancement different values of m are shown in Table 6. Table 6 clearly shows the
In order to determine the best value of m for optimal image effect of different values of m on image quality. When m = 0.5, the
brightness, entropy and average gradient were used as objective eva- entropy and the average gradient reach the optimal values, indicating
luation indices. Table 4 shows the linear enhancement evaluation re- exponential transformation has its relatively highest effect when the
sults with different values of c. As shown in Table 4, Entropy and exponential parameter m is equal to 0.5.
average gradient remain essentially the same or slightly decrease with As mentioned above, the linear enhancement method and two
the increase of c. Therefore, the linear enhancement method is not nonlinear enhancement methods were evaluated objectively. Since the
suitable for this application. linear enhancement method is not suitable for this application, the
optimal enhancement method should be selected between the two
3.3.2. Objective evaluation of logarithmic transformation enhancement nonlinear methods.
Values of entropy and average gradient under different base num- Based on the two nonlinear enhancement methods with the best
bers (1 + r) are shown in Table 5. It can be seen from Table 5 that the parameters, Table 7 presents the objective evaluation results of the
different values of (1 + r) affect the image quality. When (1 + r) = 11, initial enhanced image shown in Fig. 3(b). The comparison results
entropy and average gradient achieve the optimal values. Therefore, 11 showed that logarithmic enhancement method was more effective, so

204
J. Zhang et al. Computers and Electronics in Agriculture 151 (2018) 196–206

(a) Original image (b) Dehazed image with enhancement

Fig. 15. Comparison between original and dehazed images.

Table 8 4. Conclusion
Evaluation results of original and dehazed images.
Image Entropy Average gradient In this study, the DCP-based dehazing method was first used to
dehaze the airborne images, and then the logarithmic enhancement
Original image 6.359 6.513 method was used to enhance the dehazed images. The main conclusions
Dehazed image 6.668 11.305
are as follows:

(1) To solve the problems caused by clouds and chaotic media in re-
this method was selected as the final image enhancement method.
mote sensing images, the dark channel of the image was obtained
based on DCP theory. On the basis of this, the atmospheric light A
and transmission factor t(x) were obtained to dehaze images ac-
3.3.4. Overall evaluation of dehazed image mosaics
cording to an atmospheric scattering model.
Fig. 13 shows four image mosaics generated from the 50 original
(2) To address the brightness issue in initial dehazed images, linear
images using four image stitching methods in Adobe Photoshop: auto-
enhancement, logarithmic enhancement and exponential enhance-
matic, perspective, cylindrical and spherical. It can be seen from Fig. 13
ment were applied to enhance airborne images. The logarithmic
that the effect of automatic stitching and perspective stitching are not
enhancement method was finally selected as the optimal brightness
very good. The boundaries of the composite images appear distorted.
enhancement method based on different evaluation indices. The
From the imaging flight path, the final mosaicked image should be close
base number (1 + r) = 11 is the optimal base parameter for the
to a rectangle. Therefore, the image mosaic using spherical stitching (d)
logarithmic transformation enhancement method. Results showed
is more in line with the desired result.
that the method can effectively address the problem of luminance
Fig. 14 shows the mosaicked images after dehazing using the four
distortion of the DCP algorithm.
stitching methods, and the spherical stitching method also has the best
(3) Four indices including MSE, PSNR, average gradient, and program
result.
running time were used to evaluate the quality of the dehazed
The mosaicked images before and after dehazing do not have ex-
images, and the best dehazing parameters were obtained. Using the
actly the same area. In order to ensure the accuracy of the evaluation
logarithmic enhancement method, image enhancement achieved
results, images extracted from a common region in the mosaicked
the best effect when the dark channel window size Ω(x) is 5, at-
images based on spherical stitching as shown in Fig. 15 were used for
mospheric light A is 215/255, and the lower limit t0 of transmission
evaluation.
factor t(x) is 0.1.
Since MSE and PSNR require a reference image, entropy and the
(4) To validate the effectiveness of the dehazing method, mosaicked
average gradient were used as the evaluation indices. The evaluation
images generated from 50 airborne images before and after de-
results are shown in Table 8.
hazing were assessed. The entropy after dehazing is 6.668, 4.86%
Values of entropy and the average gradient showed that the mo-
higher than that for the mosaicked image without dehazing, and the
saicked image after dehazing was superior to the original image mosaic.
average gradient is 11.305, 73.58% higher than that for the original
For the original image mosaic, the entropy is 6.359 and the average
mosaicked image. These results demonstrate that the dehazing
gradient is 6.513. In comparison, the entropy after dehazing is 6.668,
process improves the clarity of remote sensing images and that the
which is 4.86% higher than that for the mosaicked image without de-
proposed image enhancement method is effective and can be used
hazing, and the average gradient is 11.305, which is 73.58% higher
for image dehazing in remote sensing.
than that for the original mosaicked image. The increase in entropy
shows the improvement of image quality by using image dehazing
References
technology and the increase in the average gradient shows that image
dehazing improves the clarity of remote sensing images. Therefore, the
Riaz, I., Yu, T., Rehman, Y., Rehman, Y., Shin, H., 2016. Single image dehazing via re-
proposed image dehazing method in this research is effective. liability guided fusion. J. Vis. Commun. Image Represent. 40 (PA), 85–97.
Shi, Z., Long, J., Tang, W., Zhang, C., 2014. Single image dehazing in inhomogeneous
atmosphere. Optik – Int. J. Light Electron Opt. 125 (15), 3868–3875.
Ansia, S., Aswathy, A.L., 2015. Single image haze removal using white balancing and

205
J. Zhang et al. Computers and Electronics in Agriculture 151 (2018) 196–206

saliency map. Procedia Comput. Sci. 46, 12–19. contrast-enhancement techniques: global and local area histogram equalization. Eur.
Iqbal, S.M., Abraham, A., Nizar, B.K.M., 2016. Weighted guided Gaussian single image J. Nucl. Med. 6 (6), 261–264.
dehazing. Procedia Technol. 25, 293–301. Choi, H.H., Yun, B.J., 2011. Contrast enhancement for the captured image by estimating
Ni, W., Gao, X., Wang, Y., 2016. Single satellite image dehazing via linear intensity local illumination. Opt. Rev. 18 (5), 389–393.
transformation and local property analysis. Neurocomputing 175 (Part 6), 25–39. Guo, Q., Xue, L., Tang, R., Guo, L., 2017. Underwater image enhancement based on the
Luan, Z., Shang, Y., Zhou, X., Zhao, Z., Guo, G., 2017. Fast single image dehazing based on dark channel prior and attenuation compensation. J. Ocean Univ. China 16 (5),
a regression model. Neurocomputing 245 (C), 10–22. 757–765.
Xiao, C., Gan, J., 2012. Fast image dehazing using guided joint bilateral filter. Vis. Yang, C., Hoffmann, W.C., 2015. A low-cost single-camera imaging system for aerial
Comput. 28 (6–8), 713–721. applicators. J. Appl. Remote Sens. 9 (1), 096064.
Ge, G., Wei, Z., Zhao, J., 2015. Fast single-image dehazing using linear transformation. Xie, C.H., Qiao, W.W., Liu, Z., Ying, W.H., 2017. Single image dehazing using kernel
Optik – Int. J. Light Electron Opt. 126 (21), 3245–3252. regression model and dark channel prior. SIViP 11 (4), 705–712.
Dhal, K.G., Das, S., 2017. Combination of histogram segmentation and modification to Ding, M., Tong, R.F., 2013. Efficient dark channel based image dehazing using quadtrees.
preserve the original brightness of the images. Pattern Recognit. Image Anal. 27 (2), Sci. China Inform. Sci. 56 (9), 1–9.
200–212. Russo, F., 2004. Piecewise linear model-based image enhancement. Eurasip J. Adv. Signal
Badalyan, V.G., Bazulin, E.G., 2003. Use of homomorphic filtering for upgrading flaw Process. 2004 (12), 1–9.
imaging in the expert examination of welded joints of atomic power plant pipelines. Tong, Y., Chen, J., 2017. Non-linear adaptive image enhancement in wireless sensor
Russ. J. Nondestr. Test. 39 (4), 266–272. networks based on non-subsampled shearlet transform. Eurasip J. Wireless Commun.
Wu, Y., Du, P., Shi, P., 2004. Research on wavelet-based algorithm for image contrast Netw. 2017 (1), 46.
enhancement. Wuhan Univ. J. Nat. Sci. 9 (1), 46–50. Khan, T.M., Khan, M.A.U., Kong, Y., Kittaneh, O., 2016. Stopping criterion for linear
Qureshi, M.A., Deriche, M., 2016. A new wavelet based efficient image compression al- anisotropic image diffusion: a fingerprint image enhancement case. Eurasip J. Image
gorithm using compressive sensing. Multimedia Tools Appl. 75 (12), 6737–6754. Video Process. 2016 (1), 6.
Park, S., Moon, B., Ko, S., 2017. Low-light image restoration using bright channel prior- Navarro, L., Deng, G., Courbebaisse, G., 2013. The symmetric logarithmic image pro-
based variational Retinex model. Eurasip J. Image Video Process. 2017 (1), 44. cessing model. Digital Signal Process. 23 (5), 1337–1343.
Wang, J.D., Zhang, W.T., Wang, Z.R., Xu, L.H., 2013. A fast aerial image dehazing al- Chen, S., Yao, L., Chen, B., 2016. A parameterized logarithmic image processing method
gorithm. J. Aeronaut. Eng. 34 (03), 636–643. with Laplacian of Gaussian filtering for lung nodule enhancement in chest radio-
He, K., Sun, J., Tang, X., 2011. Single image haze removal using dark channel prior. IEEE graphs. Med. Biol. Eng. Compu. 54 (11), 1793.
Trans. Pattern Anal. Mach. Intell. 33 (12), 2341. Wang, J.B., He, N., Zhang, L.L., Lu, K., 2015. Single image dehazing with a physical model
Verdenet, J., Cardot, J.C., Baud, M., Chervet, H., Duvernoy, J., 1981. Scintigraphic image and dark channel prior. Neurocomputing 149 (PB), 718–728.

206