Draft Survey Based on Image Processing

Wang Zhang, Ying Li and Wenhai Xu

Information Science and Technology College, Dalian Maritime University, Linghai Road, Dalian, China

{zhangwang, ly_123, xuwenhai}@dlmu.edu.cn

Keywords: Draft survey, climbing robot, image processing, neural network, color image segmentation.

Abstract: A draft survey system based on digital image acquisition and processing was designed and implemented to

overcome the shortcomings in the conventional manual reading method. The method of draft image

acquisition using the climbing robot as a carrier for the network camera was proposed. While climbing

along the complex surfaces of the hull under the control of a tablet, the robot could be assigned to shoot the

draft and continuously capture high-definition images covering several wave periods. The images were first

preprocessed morphologically, and the draft character was then identified by the neural network algorithm

so as to numerically represent the draft. Meanwhile, the draft line was identified by using the color image

segmentation algorithm. By doing so, the interference of fake waterline caused by the infiltration of waves

was successively eliminated, so as to accurately locate the draft line. Field experiments showed that

determining the ship’s draft by comparing the relative locations of the draft line on the numerically

represented draft could achieve a final identification accuracy of 1mm for the detected values, significantly

higher than that of 5mm by manual reading.

1 INTRODUCTION

Draft survey is an internationally accepted approach

for the cargo transfer of water transport. It is mainly

used for weighting bulk cargoes, such as coals,

which are at low price and difficult to weight by

instruments. Based on the Archimedes’ principle,

draft survey takes the ship itself as the measuring

tool and calculates the hull displacement by

measuring the ship draft. Then the cargo weight can

be calculated by combining the parameters, such as

level of ballast water and port water density, and

referring to the “Hydrostatic Data Table” (Liu,

Zhang, Sun and Yin, 2014). Conventionally, draft

survey is performed by reading the ship’s draft on

the draft marks, which usually needs to be done by

professionals. Since the weather conditions at the

arrival of a ship may not be constant, and the draft

survey must be conducted immediately after arrival

to ensure the subsequent loading and unloading

operations. The results are susceptible to wind,

waves and subjective factors. It is also impossible

for later inspection.

In this paper, to address the problems of manual

reading, a draft survey system based on image

acquisition and processing was designed and

implemented. In the system, the draft images were

captured by the climbing robot that carried network

HD camera, and the close-up imaging of the marks

near the draft line ensures the quality of the

materials for image processing. The draft was

determined by the draft numerical representation

algorithm based on neural network and the draft line

identification algorithm based on color image

segmentation. Moreover, the integrated analysis of

the images covering several wave periods could

effectively improve the detection accuracy. At the

same time, all relevant data were recorded in the

database, guaranteeing later inspection of the

measurements and thus avoiding the possible trade-

related disputes (Li, Bao and Xu, 2012).

2 SYSTEM COMPOSITION

The draft survey system was divided into an on-site

data acquisition sub-system and a back-end data

processing sub-system, as shown in Fig. 1. The data

acquisition sub-system was composed of the

climbing robot, level meter, densitometer, and tablet

computer. Under the control of the tablet, it could

capture the draft images, measure the level of ballast

water and the port water density, and upload data to

the server via 4G network. The data processing sub-

642

Zhang, W., Li, Y. and Xu, W.

Draft Survey Based on Image Processing.

In 3rd International Conference on Electromechanical Control Technology and Transportation (ICECTT 2018), pages 642-647

ISBN: 978-989-758-312-4

system centered on the programs of draft numerical

representation, draft line identification, load

calculation, and database application running on the

server. It received the draft images sent by the tablet,

calculated the ship’s load, and returned it to the

tablet for the operator while backing up in the

database for later inspection.

Figure 1: Composition of draft survey system.

3 DRAFT IMAGE PROCESSING

A typical draft image uploaded by the data

acquisition sub-system to the data processing sub-

system is shown in Fig. 2. The operator could drop

the robot at any point from the ship’s sides and

control the robot to climb on the hull with the tablet.

According to the real-time view sent by the camera,

the acquisition of draft images could be done even

for the non-visible areas.

Figure 2: The draft image of bulk carrier An** captured

by the climbing robot.

3.1 Numerical Representation of Draft

The original draft images captured by the camera are

not suitable for the numerical representation of the

draft. This is mainly due to the fact that most of the

original images are subjected to the noises of stains

and character distortion caused by shooting angle.

What’s more, it takes longer to process the larger

images. Therefore, the original images must be first

transformed into binary ones, and statistical mapping

should be plotted vertically to crop the images. Since

the draft characters with distinctive white features

can form a clear contrast against the background, it

will be easy to choose a fixed threshold for turning

the images binary after the original images have

been transformed into grayscale. The watermark of

the camera at the fixed area should be removed from

the image and the statistical mapping should be

plotted vertically as shown in Fig. 3 (a). As can be

seen by comparison, the draft character vertically

arranged in the binary image present an obvious

stepped increase in their counts in the corresponding

mapping. On this basis, it is possible to crop the core

area where the draft is located, as shown in Fig. 3

(b).

Figure 3: Preprocessed results of the draft image: (a)

vertical statistical mapping of the binary image, (b) the

cropped core

area.

If there is no obvious noise in an image, the

mapping can be then plotted horizontally, in order to

crop image of single characters and thus start

character identification. However, before that, noise

reduction and shear correction of the image could

effectively improve the accuracy of subsequent

character identification. Given that the characters on

the standard metric draft markings contain only the

Arabic numerals “0-9” and the capital letter “M”,

this paper adopted morphology-based noise

reduction algorithm. The key operations involved

are mathematically described by Eq. (1) and (2):

(a) (b)

Data acquisition sub-

ystem

Climbing robot carrying network HD camera

Tablet Level meter Densitometer

Data processing sub-

ystem

Data server

Draft numerical representation Draft line identification

Load calculation Database application

Draft Survey Based on Image Processing

643



BBABA  ⊙

(1)



BBABA ⊙

(2)

where



denotes opening operation,  denotes

closeing operation,

⊙

denotes erosion, and



denotes dilation. As implied by the combination of

erosion and dilation, opening operation can

eliminate the small prominences of isolated noise

points, thin lines, and edges, while closeing

operation can eliminate the internal small holes and

curves (Said, Jambek and Sulaiman, 2016).

Comparing Fig.3 (b) before processing and Fig.4 (a)

after processing, it can be found that the jagged

prominences of edges of the characters are

noticeably eliminated, and the noises caused by a

few large stains can be removed in further character

cropping.

Figure 4: Preprocessed results of the draft image: (a)

image after noise reduction, (b) the cropped character, (c)

mapping of the character image in different directions, and

(d) the character image after shear correction.

The majority of the cropped characters would

still be deformed, because of the shooting angle of

the robot, as shown in Fig.4 (b). Since it could be

difficult for the robot to maintain a stable stance on

the raised draft characters, images could only be

taken from the diagonal tops of the characters. As a

result, the draft characters in an image would

inevitably be sheared horizontally. The angle of

shear-warp can be extracted by using the mapping

method shown in Fig.4 (c), that is, when the

mapping angle θ is equal to the shear-warp angle,

the mapping length L can be the minimum (Liu,

2017). Let the coordinates of a pixel point in the

image be P (x, y), after the shear correction by using

affine transformation, the coordinates P' (x', y') can

be expressed as follows:









































1100

01tan

001



(3)

The correction is performed using a bilinear

interpolation method. The output pixel value is the

average of a 2×2 field samples in the input image, as

shown in Fig.4 (d).

Among many of the draft character identification

methods that have been tried, the template matching

and the neural network algorithms are considered to

be effective (Ran and Peng, 2012). The former often

uses simple features for comparison, so the

algorithm can be conveniently implemented; but the

characters like “6”, “8”, and “9” that are

morphologically similar are poorly distinguished,

especially if the characters are defaced by sea water

erosion. The latter identifies the character by the

weights of various features, resulting in higher

identification accuracy, and its complex algorithm is

affordable for high-performance servers (Kim and

Xie, 2015).

The typical BP-neural network used in this paper

consists of three layers: input, hidden, and output

layers. Neurons in adjacent layers are connected to

each other, while neurons in the same layer are not

connected. As shown in Fig.5, the neural network’s

inputs are x

, x

, …, x

; the hidden layer’s

outputs are y

, y

, …, y

; the neural network’s

outputs are o

, o

, …, o

; the weight from

input layer to hidden layer is v

; the weight from

hidden layer to output layer is w

. For neuron

outputs at the hidden and output layers, there is:





















iijjj

xvfnetfy

(4)





















jjkkk

ywfnetfo

(5)

The commonly used unipolar Sigmoid function









is selected as the activation function.

According to the specific needs of draft character

identification, the input layer contains 93 neurons,

covering various parameters such as the character’

contour, projection, and discrete Fourier features.

The output layer contains 11 neurons, corresponding

to the numbers “0-9” and the letter “M”, respectively.

The hidden layer contains 48 neurons and connects

the input and output layers via the full connection, in

order to achieve the forward transmission of

information and the backward transmission of error.

When the output results generated by the input

(a) (b) (d)

(c)

L (θ = 0°)

L (θ = -19°)

ICECTT 2018 - 3rd International Conference on Electromechanical Control Technology and Transportation

644

feature parameters do not match the expected values,

the error will be transmitted backwards. Moreover,

the weight between layers will be adjusted in a

gradient descent manner, until the correct output

results are obtained after repeated attempts, thus

forming a widely applicable structure of neural

network.

Figure 5: The structure of neural network used for

character identification.

3.2 Identification of Draft Line

The key to draft line identification is how to

accurately identify the intersection between the hull

and the water surface. Drawing on the previous

experience of edge detection, the effects of various

edge detection operators such as Sobel, Reborts, Log,

Zerocross, Canny, and Susan were compared in the

experiment (Yu and Zhang, 2015). Among them,

operators Sobel and Reborts can digitally

approximate the first derivative to extract the edges,

but a lot of noises may be left in the processed image.

Operators Log and Zerocross use the filtering

function to carry out convolution of the image.

Compared to operators Sobel and Reborts, they are

effective in smoothing the image, but double edges

are easily generated as well. The difficulty in

applying operator Canny lies in how to determine

the size and threshold of Gaussian filter, while the

fixed Gaussian filter can hardly detect the edge

structures of different scales. Operator Susan has a

strong anti-noise capacity, but an appropriate

threshold is also difficult to select.

In addition, when using conventional edge-

detection operators for image processing, there are

often two indistinguishable draft lines, especially

when the weather is fine. As clearly shown in the

original Fig.6 (a), the upper half of character “6” has

an area created by undulating infiltration on the hull.

The gray values of this area and the non-infiltrated

area form a significant discontinuity, even more

evident than the difference between them and the

water surface. In the practical detection, therefore, it

can be tricky to determine which one is the real draft

line. Furthermore, the fluctuation of the water

surface and the resulting changes in the reflected

light may create obvious discontinuity in the gray

image, seriously affecting the actual effect of the

conventional edge detection operators.

Figure 6: RGB components and their histograms of the

draft image: (a) the draft image, (b-d) RGB components,

(e) histograms of the draft image, and (f-h) histograms of

RGB components.

Given the obvious difference in colors between

the hull and the water surface, this paper used the

color image segmentation algorithm to achieve the

accurate identification of the draft line. As suggested

by Fig.6 (e), which is the histogram of Fig.6 (a), the

bimodal distributions are caused by the subtle color

differences. By extracting the components of R (red),

G (green), and B (blue) of the

draft image and

comparing their histograms, it can found that the

histogram corresponding to the R

component presents

a significant bimodal distribution, while the other

components show the unimodal distribution. The

hulls of bulk carriers are always painted bright red,

while the water surrounding the cargo terminals is

generally muddy. As a result, there can be an evident

contrast between the two, leading to a distinctive

bimodal distribution of the histogram of R

component.

Influenced by the unimodal distribution of the other

two components, this feature is not obvious in the

original image. For this reason, the minimum gray

value between the two peaks in the histogram of R

(a) (b) (c) (d)

(e) (f) (g) (h)

Input layer

Hidden layer

Output layer

Draft Survey Based on Image Processing

645

component was taken as the threshold, so as to

accurately segment the images. The theoretical basis

is that the gray values of the pixels adjacent inside

the targets or the background area similar, while

those of the pixels between the target and the

background are different (Peng, Zhou and Lei, 2017).

Therefore, the target and the background correspond

to different peaks in the histogram. For the pixel

point

R (x, y) of R component, the valley T between

two peaks in the histogram is selected as the

threshold. Then the segmented binary image

(x,

y) can be expressed as:



















TyxRb

TyxRa

yxR

(6)

where

a=1 denotes the target, b=0 denotes the

background, that is, the hull and the water surface

are segmented in the image.

As shown in Fig.7 (a), in the segmented result

obtained from R

component, the hull and the water

surface show an obvious margin, while the fake

waterline formed by wave infiltration on the hull

does not leave evident traces. Such effect is mainly

due to the little impact generated by the infiltration

itself in the image of R

component. Meanwhile, the

valley between the two peaks on the histogram is

taken as the threshold for segmentation, which also

helps eliminate the influences caused by the small

difference in the gray values of the adjacent pixels

inside the target or background. In addition, to

improve the adaptability of the image segmentation

algorithm, the histograms of different color

components can be compared other than the

histograms of RGB

components.

Figure 7: Identified results of the draft line based on color

image segmentation, (a) binary image obtained from R

component, (b) pixels on the edges, (c) the detected draft

line, and (d) mapping of the draft line in the original

image.

The edge pixels in the image are extracted as

shown in Fig.7 (b). The details show that although

the edge features between the hull and the water

surface can be obtained by the above method, the

pixels at the edges usually do not fully characterize

the edge, especially if the draft line stretches over

the draft character. For the edge fractures due to

noises and uneven lighting, as well as the other

effects of introducing grayscale discontinuities,

Hough transform is usually used to assemble the

edge pixels into meaningful continuous segments.

The basic strategy is as follows: A set of straight

lines that pass a specific point in the image are

converted to a curve under polar coordinates, the

peaks of the curve intersections under polar

coordinates are counted in an accumulator, and then

the peak corresponds to a straight line with many

collinear points in the image (Yan and Yang, 2015).

For the identification of the draft line, given that

adjusting the climbing robot’s location and arm can

provide a better shooting angle for the HD camera,

the location of the draft line is limited within the

lower half of the image, and the angle of the draft

line is limited to ±15°. This not only facilitates

reducing the interference in the image, but also

accelerates the processing speed of Hough transform,

as shown in Fig.7 (c). Finally, the resulting line can

be remapped to the corresponding location in the

original image as the draft line, as shown in Fig.7 (d).

3.2 Calculation of Draft Value

After numerical representation of the draft and

locating of the draft line, the draft value can be

obtained immediately by comparing the relative

location of the two, but one of the details will make

a difference in the identification accuracy. Since

there is an angle between the camera and the draft,

the distances may differ between the numerically

represented draft characters. Hence, it is necessary to

determine the variation pattern through the fitting

approach, and thus the depth value represented by

the distance between the draft line and the last

character. Considering that Hough transform is used

in identifying the draft line, and that the draft line is

located by many edge pixels, a locating accuracy at

sub-pixel level could be achieved theoretically.

Accordingly, the calculation accuracy of the draft

value reaches 1mm, significantly higher than the

5mm achieved by manual reading.

4 CONCLUSIONS

Draft survey based on digital image acquisition and

processing is an innovative approach that uses

pioneering technologies to overcome the inherent

(a) (b) (c) (d)

ICECTT 2018 - 3rd International Conference on Electromechanical Control Technology and Transportation

646

disadvantages of conventional manual reading. The

climbing robot that carries the camera to capture

images makes it possible to obtain the close-up

images of the draft. The robot can climb along the

hull under the control of a tablet, and continuously

captures several HD images when approaching the

draft. The identification algorithm based on neural

network and color image segmentation ensures the

accurate identification of the draft, with a calculation

accuracy of 1mm for draft values in the image. In

particular, by combining several draft images, the

calculated results could effectively reduce the

impacts of wave undulation. Moreover, the entire

measurement process could be recorded in the

database, guaranteeing later inspection of the

measurements and thus avoiding the possible trade-

related disputes.

REFERENCES

Liu, C.L., Zhang, X.F., Sun, X.F. and Yin, Y., 2014.

Improved draft survey method based on densified

table of offsets. Journal of Traffic and Transportation

Engineering, 14(3): pp.58-64.

Li, H.X., Bao, Y.F. and Xu, Z. Y., 2012. The research on

the risk evaluation and management system for draft

survey. Journal of Inspection and Quarantine, 22(4):

pp.4-7.

Said, K.A.M., Jambek, A.B. and Sulaiman, N., 2016. A

study of image processing using morphological

opening and closing processes. International Journal of

Control Theory & Applications, 9(31): pp.15-21.

Liu, L.L., 2017. Research on image segmentation

technology for a license plate recognition. Bulletin of

Science and Technology, 33(4): pp.125-129.

Ran, X. and Peng, J.H., 2012. Ship draft mark recognition

based on Image Processing. Journal of Shanghai

Maritime University, 33(2): pp.6-9.

Kim, I.J. and Xie, X.H., 2015. Handwritten Hangul

recognition using deep convolutional neural networks.

International Journal on Document Analysis and

Recognition, 8(1): pp.1-13.

Yu, J.D. and Zhang, X.M., 2015. Edge detection algorithm

for lines on microscopic image. Optics and Precision

Engineering, 23(1): pp.271-281.

Peng, H.，Zhou, X.Z. and Lei, Y.J., 2017. Color image

segmentation based on prior color covariance

constraint. Computer Engineering, 43(4): pp.251-256.

Yan, H.R. and Yang, M.S., 2015. Line extraction based on

improved Hough transform. Infrared Technology,

37(11): pp.970-975.

Draft Survey Based on Image Processing

647