Computers in Industry 92 (2017) 152160

Contents lists available at ScienceDirect

Computers in Industry
journal homepage: www.elsevier.com/locate/compind

A ne grained digital textile printing system based on image

registration
Jinxiong Rena , Gang Chena,* , Xiaoyan Lib
a
College of Computer Science and Technology, Zhejiang University, PR China
b
Department of Multimedia, Netease (Hangzhou), Inc., PR China

A R T I C L E I N F O A B S T R A C T

Article history:
Received 12 October 2016 Printed fabrics differ from non-planar fabrics (e.g., embroidery) in their rich colours, various patterns and
Received in revised form 15 May 2017 low costs. However, non-planar fabrics are considered more attractive, expensive in appearance and
Accepted 9 August 2017 luxurious owing to their solid texture patterns, oats and luster contrasts. To produce fabrics that are
Available online xxx both colorful and lavish and also have color gradients, this paper proposes a novel ne-grained digital
printing system that combines digital textile printing and non-planar fabrics. Note that because
Keywords: distortion is one of the characteristics of the fabric, directly printing an ink-design pattern onto non-
Textile manufacturing planar fabric gives poor performance. This work develops a solution to accurately align the ink-design
Fine grained digital printing system
pattern with the fabric, consisting of both hardware and software parts that are used to take pictures of
Printed fabrics
the fabric, employ image registration-based computer vision technology to deform the ink-design
Non-planar fabrics
Machine vision pattern according to the deformability of the fabric, and print the deformed ink-design pattern onto the
Image registration fabric. The system is exible and highly automated, and it makes mass production possible. In the
prototype, ne quality fabrics are produced efciently (a fabric of 0.12 m2 in less than 4 s).
2017 Elsevier B.V. All rights reserved.

1. Introduction using hoops have been proposed to ensure accurate alignment

Digital textile printing is a new technology to make ink-designs
on textiles, which began in the late 1990s to replace traditional
analogue screen-printing [13]. Compared with analogue screen-
printing, digital textile printing is simpler and generates fewer
emissions because it skips the processes of color separation,
drawing and screen preparation [4]. Digital textile printing makes
it possible to create fabrics with vibrant colors, photo-quality
images and personalized designs [5]. However, textile printing
products are considered low-end in comparison with non-planar
fabrics (e.g., embroidery, lace) due to the latter's grains and luster
contrasts. The problem is that non-planar fabric products are
limited in terms of colors and efciency.
To produce colorful and lavish fabrics with solid texture
patterns, approaches to combining non-planar fabrics and printing
have been studied for over two decades. The tremendous challenge
is how to achieve a perfect alignment of fabric with ink-design,
considering the deformability of fabric. Apparatus and methods
* Corresponding author at: College of Computer Science and Technology,
CaoGuangBiao Building, Zhejiang University, 38 Zheda Road, Hangzhou 310018,
China.
E-mail address: cg@zju.edu.cn (G. Chen).
between the screen-printing and automatic embroidery machines
[6,7]. Kuki et al. [8] designed embroidery sewing machines that
allow both embroidery and print images to be formed on the
workpiece cloth.
However, there are two main drawbacks to the existing
approaches. First, specic apparatus are proposed for embroidery,
which cannot be generalized to different fabrics. Second, these
methods require signicant human intervention.
In contrast to previous works, this paper presents a novel ne-
grained digital printing system based on image registration
technology [9,10] that produces digital ink-designs on non-planar
fabrics. This solution consists of specic hardware and suitable
algorithms to overcome the deformability of fabric (as shown in
Fig. 1). The key hardware components in the proposed printing
system include an industrial camera, an LED, a computer and a
digital textile printer. In this paper, algorithms are proposed to
align the ink-design pattern with the fabric pattern. The basic idea
of this solution is to take pictures of the fabric, employ image
registration-based computer vision technology to deform the ink-
design pattern according to the deformability of the fabric, and
print the deformed ink-design pattern onto the fabric. Using the
proposed system, customized fabrics on which the fabric pattern
and the ink-design pattern are perfectly matched can be obtained.

http://dx.doi.org/10.1016/j.compind.2017.08.003
0166-3615/ 2017 Elsevier B.V. All rights reserved.
 J. Ren et al. / Computers in Industry 92 (2017) 152160
Fig. 1. An example shows that direct printing cannot achieve satisfying results. The
ink-design pattern does not align perfectly with the stitch contours.
The proposed system is more exible, highly automated and makes
mass production possible.
Regarding the proposed printing system, there are several key
challenges to be addressed, including a new processing ow to
produce customized fabrics and several algorithms to align the
fabric pattern with the ink-design pattern in real time.1 As far as
known, the proposed system is the rst to combine digital textile
printing and non-planar fabrics precisely to produce new
customized fabrics.
The main contributions can be summarized as follows:
(i) A novel ne-grained digital printing system which combines
digital textile printing and non-planar fabrics to produce new
customized fabrics.
(ii) A new processing ow to produce customized fabrics on
which fabric patterns and ink-design patterns are perfectly
matching.
(iii) Several algorithms to perfectly align non-planar fabrics (e.g.,
embroidery, lace, etc.) with ink-design patterns.
2. Related work
Textile printing originates from ancient China as the process of
printing color to textiles in designs [11]. With the development of
digital technology, digital textile printing emerged in the 1990s
[12,13]. It includes two methods for digital textile printing: indirect
heat transfer printing and direct ink-jet printing [14].
Approaches of combining non-planar fabrics and printing have
been studied for over two decades. Allen [6] proposed a
combination embroidery/screen printing apparatus and method
are provided to substantially concurrently reproduce a pattern on
textile substrate wherein a rst pattern portion of each such
reproduction is screen printed and a second portion of each such
reproduction is embroidered. This work employed mechanic
apparatus to achieve alignment between screen printing and
embroidery. However, much human intervention is needed in
manufacturing. Buck [7] presented one embodiment of a system,
method, and apparatus for screen printing and embroidering a
garment mounted in a hoop. Kuki et al. [8] suggested that
embroidery sewing machines allow both embroidery and print
images to be formed on the workpiece cloth. Mass manufacture is
impossible for these methods because of complicated operations.
In contrast to previous approaches, the proposed system which
combines digital textile printing and non-planar fabrics is more
exible, highly automated and makes mass production possible.
This solution is composed of specic hardware and algorithms on
top it, which performs image acquisition, distortion detection,
image alignment and prints deformed ink-design pattern. The
1
Real time means the image whose size is about 300
6000 pixels was dealt
with in 4 s in which the fabric moves from imaging zone to printing zone.
153
alignment between the fabric image and the ink-design image is
addressed by a feature-based image registration method [1517].
Because fabrics show different characteristics, various algorithms
are employed to handle them.
3. Fine-grained digital printing system
3.1. Overview
Digital textile printing technology is becoming increasingly
popular for textile printing in modern industry. It produces fabric
products with vibrant colors, photo-quality images and personal-
ized designs. This paper proposes a ne-grained digital printing
system that combines digital textile printing and non-planar
fabrics to produce new colorful and high-grade fabrics. Note that
distortion is one of the characteristics of the fabric. To perfectly
align the ink-design pattern with the fabric pattern, machine vision
technology is employed. The system is a combination of hardware
and software approaches. In the following sections, the system will
be introduced in detail.
3.2. Approach
This work relies on machine vision to achieve perfect alignment
of the ink-design pattern and the fabric pattern through registering
their images. Normally, the fabric pattern and the ink-design
pattern contain repeating patterns. A single fabric pattern repeat is
referred to as a fab-design.
The approach is as follows. The live fabric image is acquired by a
monochrome CCD camera in the imaging zone (Fig. 2) while the
fabric is transported on a black conveyer and is then matched to a
prototype image (a single ink-design pattern repeat). During this
matching, the system determines the fabric distortions (compared
to the single ink-design pattern repeat) in terms of the matching
result, and rectications are performed on the ink-design pattern
to compensate for these distortions in real-time. After these
corrections are performed, the deformed ink-design pattern is
printed onto the fabric in the printing zone.
There are four steps in our approach: system initialization,
image acquisition, image processing, and printing, which will be
introduced in detail in section processing ow.
3.3. Processing ow
Once the fabric and the ink-design pattern to be printed on it are
picked up, the live fabric image is processed by the proposed
approach. In this section, the processing ow consisting of four
steps is introduced as shown in Fig. 3. For each fabric, step 1 is
performed only once at the beginning. Then, all live fabric images
are processed by step 2step 4.
Fig. 2. Side view of the ne-grained digital printing system. (1) The industrial
camera and LED. (2) Imaging zone. (3) Printing head. (4) Printing zone. (5) Conveyer.
154 J. Ren et al. / Computers in Industry 92 (2017) 152160

Fig. 3. Processing ow.

1. System Initialization. This step is performed to nd the start
position. At the very beginning of the production process, the
position of the fabric under the camera with respect to the
prototype image is determined manually. Once the start point
has been determined, the live image is synchronized with the
prototype image. This step is performed only once at the
beginning for each fabric.
2. Image acquisition. Image acquisition is one of the most
important elements in a machine vision system [1820]. A
monochrome CCD camera is used in this system because most of
the fabrics used are monochromatic, and this camera has high
resolution to capture more details of the fabric. During the
production process, the fabric to be printed is transported over a
black-lit stage, as shown in Fig. 2. When the fabric proceeds into
the imaging zone, the camera yields high-resolution images.
3. Image processing. As previously stated, the acquired image is in
grey scale and thus is represented by a matrix whose elements
are the brightness intensity level corresponding to an image
pixel. Once acquired, the live fab-design image is coarsely
matched to the prototype image according to the size of the
repeat pattern contained in the fabric, since the live image is
synchronized with the prototype image in step 1. The proposed
algorithms are employed to detect the fabric distortions and
compute a transformation matrix to correct the ink-design
pattern so that the deformed ink-design pattern is perfectly
aligned with the live image.
4. Printing. The corrected ink-design is printed onto the fabric.
obtain real-time high-resolution images, a line-scan camera is
selected.
The line-scan camera used in the system is Dalsa SF-10-02K40
whose resolution is 2048 pixels and the line scan rate is 18 kHz.
In this system, the image area captured by the cameras in one
motion is called one Pass. The fabric images are handled Pass by
Pass.
LED. Since illumination is crucial to image acquisition, an LED
light moves together with the camera to ensure the same
illumination conditions for all images.
4. Registration algorithms
4.1. Overview
As is mentioned before, it is crucial to align the ink-design
pattern with the fabric when deformability occurs. This issue is
addressed by being viewed as an image registration problem.
Fabric image is the reference image, and ink-design image is the
target image.
However, it is impossible to align the fabric image with the ink-
design image directly because they are in different color spaces and
there is no correspondence between their patterns. As shown in
Fig. 4, the fabric image is a gray image, while the ink-design image
is in RGB color space. Hence, an intermediate image between them
is introduced. This intermediate image is a repeat fab-design
pattern image manually obtained from early Pass of the fabric

3.4. System conguration

Fig. 2 shows the printing system, which consists of an online

vision system and a digital textile printer. Its main components are
a monochrome CCD camera, an LED and a computer with a GPU.
Camera. A monochrome CCD camera is employed to capture
the live fabric image because most of the fabrics used are
monochromatic, and this camera has high resolution and speed.
Two solutions are proposed to capture the fabric in the system.
The rst solution is that at a distance of 50 cm from the printing
zone, eight area array CCD cameras are placed above the conveyer,
as shown in Fig. 2. The reason is that the width of the fabric is too
large to be captured by a single camera, considering its eld of view
and the distance between the camera and the fabric. The cameras
are synchronized to capture images at the same time when the
conveyer stops. Then, multiple images with overlapped elds of
view are combined to produce a panoramic image of the fabric.
However, this process is non-trivial, since it is difcult to align
cameras and stitch images from different cameras in real time.
The second solution is that a line-scan camera is placed in the
same position mentioned above. Line-scan cameras are widely
used in industrial, medical and image processing elds. This type of
camera captures images of rapidly moving objects or materials,
such as rolls of paper or sheet metal. In this system, when taking
pictures of the fabric, the camera moves automatically from one
side to another to scan the fabric instead of moving the fabric. To Fig. 4. Fabric pattern and ink-design pattern images are in different color spaces
and there is no correspondence between their patterns. Fabric image is a gray
image, while ink-design image is in RGB color space.
 J. Ren et al. / Computers in Industry 92 (2017) 152160
Fig. 5. Five-stage algorithm ow. Given a instance of the repeat fab-design image in the live fabric image, the approach is designed to align the instance image with the
selected intermediate image. Based on the correspondence between the intermediate image and the ink-design pattern, transformation is performed to the ink-design
pattern in terms of the alignment result and the deformed ink-design image is printed onto the fabric.
image. The intermediate image and the run-time fabric image are
actually aligned because they have similar pattern and color space.
The algorithm transforms this registration problem into a ve-
stage process as shown in Fig. 5:
1. (1) Select one repeat fab-design pattern image of the fabric as
intermediate image.
(2) Get a binary representation of the intermediate image. In
the binary image, the foundation fabric is separated from
fab-design patterns.
(3) Select control points in binary image and extract features.
2. Manually generate correlations between intermediate image
and ink-design image.
3. Acquire live fabric images, and use the image segmentation
technique to get their binary images.
4. For every control point in the intermediate image, its best match
is searched in the fabric binary image.
5. Compute the transformation matrix to deform the ink-design
image.
4.2. Types of fabrics
Since fabrics show different characteristics, they are handled
separately. There are three types of fabrics this system can handle:
(1) embroidery with only stitches along shape; (2) embroidery
with stitches inside shape; and (3) lace.
Embroidery with only stitches along shape is created only by
stitches along its pattern shape. It is easy to separate the
embroidery pattern and the foundation fabric with basic image
processing operations because the stitches and the patterns of this
kind of embroidery are simple as shown in Fig. 6.
Embroidery with stitches inside the shape has inherent solid
texture patterns and thickness shadows, which makes it difcult to
lter and complete region separation. The random forest [21]
based texture classication2 is employed to address light variation
on embroidery surface and noised texture structure so as to nish
2
This code is implemented by our own as described in [21].
155
region separation. Finally, image matching is used to align ink-
design patterns with embroidery patterns.
Lace is made of thread in a weblike pattern. Its pattern shape is
not captured by contour but by threads inside it. Therefore, lace is
handled with the same processing ow as the embroidery with
stitches inside the shape.
4.2.1. Embroidery with only stitches along the shape
In the proposed algorithm, the pattern of this kind of
embroidery is treated as a point set and it is assumed that the
shape is essentially captured by a nite subset of points.
Specically, its pattern is represented by a discrete set of points
sampled from contours. Fig. 6b shows sample points for one shape.
Our algorithm consists of different modules: segmentation, feature
exaction, searching and transformation.
Segmentation. To acquire the binary representation of the live
fabric image, the noises are eliminated and the shape contour is
sharpened by adopting median and Laplacian lters. Then the Otsu
image threshold algorithm is applied [22,23]. An example is shown
in Fig. 6.
Feature exaction. A 60-dimensional feature vector for every
control point using local shape context is extracted. The basic idea
of the shape context descriptor for a point is to extract a histogram,
which assigns the other points of the shape into a log-polar
histogram, thereby capturing the object shape [24]. For each
control point in this case, 5 radial and 12 angular bins are used to
capture only local points instead of all the points of the shape. The
textile pattern in one Pass cannot be fully identied (as shown in
Fig. 7) so that the local points are used as features too.
Search. For every control point in the intermediate image, its
closest point in binary run-time fabric image is searched. A
searching region (see Fig. 7b, dashed rectangle) is generated
according to the fabric information such as pattern distance and
pattern size. Points in this region are viewed as candidate points
and local shape context as features are extracted for them. Finally,
the closest point is obtained using the equation as follows:
X
60
Pji argmax Pi k 6 0; Q j k 6 0 1
k1
156 J. Ren et al. / Computers in Industry 92 (2017) 152160

Fig. 6. Fabric image and its segmentation result.

Q indicates the set of control points in intermediate image. [] is the in this section. The classication techniques of texture are grouped
Iverson bracket. For control point Qj in the intermediate image, Pji is in several categories, (1) statistically, (2) signal processing and (3)
searched in the searching region. Pi(k) indicates the kth element of structural. The former two are the most widely used because they
feature vector Pi. The number of elements where P and Q are not 0 can be directly applied to all types of texture. For example, local
simultaneously is counted. binary patterns (LBP) [26,27] and Gabor lters [28], wavelets and
independent component analysis (ICA) are popular feature
To ensure matching performance, point Pji can be pruned if its
extraction methods [29]. In this paper, Gabor lters are used to
matching score is below a given threshold. Meanwhile, the printing
extract features. Instead of generating a binary image for the run-
system is suspended, waiting for manually operations.
time fabric, a random forest classication model [21] is employed
Transform. After gaining the closest points for all the control
to create a probability map which shows the certainty of every
points, the non-rigid deformation is performed to the ink-design
pixel being embroidery pattern. For full-embroidery, the probabil-
pattern. To deform the image in real time, a strategy is proposed.
ity map and binary representation of intermediate image (Fig. 8)
Firstly, given the corresponding points, thin plate spline (TPS) is
align with each other.
used to estimate warping parameters [25]. Then, an uniform grid of
Model training. To get the probability map of fabric image, a
points (every 1 cm) is generated in the ink-design image and
random forest classier which feeds feature of every pixel is
triangulate the ink-design image with these grid points. Instead of
trained. Random forests are a combination of tree predictors which
deforming the whole ink-design image with TPS, only grid points
are tted on various sub-samples of the dataset. It corrects a single
are transformed by TPS and points in every triangle are applied
tree's habit of overtting. In this paper, a discrete set of Gabor
afne transformation dened by three vertices (grid points). In this
lters consisting of 5 scales (spatial frequency) and 8 distinct
way the inherent parallelism of the strategy is exploited to speed
orientations is adopted to extract texture features. Gabor lter is
up the ink-design pattern deformation.
widely used for texture classication. It is similar to human visual
system and found to be particularly effective for texture
4.2.2. Embroidery with stitches inside the shape
representation and discrimination. The training process is as
In this section, the algorithm for embroidery with stitches
follows. In Step 1 as shown in Fig. 5, repeat fab-design pattern
inside its shape is introduced. This kind of fabric is referred as full-
image of the fabric is selected as the intermediate image and get its
embroidery.
binary image manually. Then, a 40-dimensional texture feature
The difference between the embroidery pattern and the
vector is extracted for every pixel in the intermediate image and
foundation fabric is the texture. Texture classication is employed

Fig. 7. Red points are control points and their best matches. As shown, taking whole contour points as feature is impossible. (For interpretation of the references to color in
this gure legend, the reader is referred to the web version of this article.)
 J. Ren et al. / Computers in Industry 92 (2017) 152160 157

Fig. 8. Intermediate image and its binary representation.

values of the binary image are viewed as pixel labels. Finally, a
random forest classication model is trained.
Feature exaction. For each control point in the binary
representation of the intermediate image, image patch around it
is selected as its feature. For a candidate point in the search region,
it is a probability patch centered around it.
Search. For a test fabric image, its probability map is generated
by the model trained before. For each control point in the
intermediate image, a search region is generated according to the
fabric information to search for the closest point in the probability
map. An example is shown in Fig. 9.
To improve the accuracy of matching, a multi-scale matching
algorithm is proposed. For control point M, the feature distance is
dened as
X
S
M0best argmin jPs M  P0s M0 j
s1
where S is the number of scales, M denotes the control point, M0 is
the candidate point, Ps(M) is feature patch centered around point
M at scale s and P0s M0 is the probability map patch centered
around candidate point M0 at scale s. The closest point found by the
multi-scale sum of absolute differences (SAD) loss function shows
good results. SAD measure is less time consuming, which is
suitable for this application.
GPUs. In the proposed system, operations like Gabor feature
extraction, random forest training and predicting are considerably
time consuming and computationally expensive. To overcome the
problems, GPUs are considered to speed up the computation of
registration process. Cuda random forest and GPU-version of Gabor
lters are employed in the printing system. When running with an
Nvidia Telsa K40M, for an image with pixel size 7000
400, Gabor
feature extraction and random forest prediction take about 1.9 s
and 0.5 s, which are about 15 times and 26.5 times faster,
respectively.
Transform. This operation is the same as that for embroidery
with only stitches along the shape.
5. Results
In this section, experimental results of the proposed algorithms
are presented for fabrics, embroidery, full-embroidery and lace.
The algorithms are implemented in C++ and CUDA. The completely
innovative new fabrics produced in factory are shown.
There are two main parameters in registration problem:
number of control points (denoted by n) and feature parameter.
Control points generation. For all fabrics, one control point
2 every 0.5 cm is selected to reduce the computation overhead. The
candidate control points are detected by corner detector and the
redundant ones are removed manually.
Feature parameter. For embroidery, the parameters of local
shape context are set as 5 radial and 12 angular bins. The patch to
extract features is 30
30 pixels. The searching region is 60
60
pixels. Result in 1st row of Fig. 10 shows great alignment of the
proposed algorithms.
For full-embroidery, features are extracted by a discrete set of
Gabor lters consisting of 5 scales and 8 distinct orientations. The
computation cost is less than 2 s. During matching process, the
closest points are searched, at three scales, 15
15 pixels, 30
30
pixels, 45
45 pixels. Example results are shown in 2nd and 3rd
row of Fig. 10. The numbers of trees (denoted by nt) in the test are
set 10, 20 and 100. If nt decreases, the predicting time is also
reduced from several seconds to less than 0.5 s with little
performance degradation (Fig. 11).

Fig. 9. Test image and its probability map. The probability map and the binary image of the intermediate image are actually registered.
158 J. Ren et al. / Computers in Industry 92 (2017) 152160

Fig. 10. The 1st row is result on embroidery. The 2nd and 3rd row are results on full-embroidery.

In Fig. 12, result on lace is shown. The same processing ow of and background lies in threads density and the ways the stitches
embroidery is used to deal with lace. Lace is made of yarn in netlike are put. It is assumed that lace pattern shape is not captured by
pattern. The lace pattern and foundation fabric are made at the contour but threads inside it. In the case of lace, torque distribution
same time unlike embroidery. The difference between lace pattern based on torque of patch [30] is used as features because perfect

Fig. 11. Probability maps obtained by different random forest model (number of trees).
 J. Ren et al. / Computers in Industry 92 (2017) 152160 159

Fig. 12. An example of lace.

Table 1
The mean values and SDs of Hausdorff distance of the two aligned contours in pixels and millimeters. The rst row of every method shows results in pixels and the second row
shows results in millimeters.

Fabrics Embroidery Full-embroidery 1 Full-embroidery 2 Lace

Our method 4.3137  0.5 4.5137  0.3 4.5549  0.3 4.6859  0.2
0.9131  0.11 0.9554  0.06 0.9641  0.06 0.9918  0.04

RC 4.8958  1.1 4.9375  1.3 4.9737  1.3 5.1346  1.4

1.0363  0.23 1.0451  0.28 1.0528  0.28 1.0869  0.30

MI 5.1268  1.0 5.3686  0.9 5.3974  0.9 5.7993  1.2

1.0852  0.21 1.1364  0.19 1.1425  0.19 1.2276  0.25

pattern contour could not be obtained by edge detection method The segmentations are manually obtained and treated as ground-
(owing to noise). For point o, its feature is dened by the torque truth. The proposed method is compared with the minimization of
distribution in patch P (centered on o), a two-dimensional matrix residual complexity (RC) [31] and local mutual information (MI)
M of which element Mij is the torque of patch centered on point at [32].
position (i, j) in patch P. Feature patches used in the experiments is The evaluation method is based on the Hausdorff distance of the
at three scales, 15
15 pixels, 30
30 pixels, 45
45 pixels. two contours of the reference image and transformed target image
[3335]. The Hausdorff distance [33] is dened as
5.1. Registration accuracy
HA; B maxfhA; B; hB; Ag 3
In this section, the registration accuracy of the proposed
methods has been tested upon real images. For every fabric in
hA; B max f min fda; bgg 4
Figs. 10 and 12, 100 live fabric images are selected as test images. a2A b2B

Fig. 13. New fabric products.

160 J. Ren et al. / Computers in Industry 92 (2017) 152160
where A and B denote the point sets of the contours in two images.
Table 1 shows the average values of Hausdorff distance H and the
standard deviations of the four fabrics obtained by different
methods. As listed in Table 1, the proposed method gives low
values of Hausdorff distance H and low standard deviations for the
registrations of the real fabric images. In the proposed system, the
resolution of the images is 120 dpi. Therefore, the printing error is
about 1 mm. And the running time is less than 4 s for one Pass. One
pass is 8
150 cm, namely an image whose size is about
300
7000 pixels. Thus, the productivity of the printer satises
practical production need.
Fig. 13 shows examples of fabrics produced by the proposed
system. Fabrics are white backgrounds with colorful ink-design
printed on them. The printing system makes fabric more
expressive, attractive and luxury. Digital textile printer from
Atexco is used in the experiments.
6. Conclusions
In this paper, a novel ne-grained digital printing system based
on machine vision technology is presented which combines the
digital textile printing and non-planar fabrics such as embroidery
and lace. The approach is capable of printing the ink-design pattern
onto non-planar fabrics. This system consists of a linear array CCD
camera, an LED, a computer and a digital textile printer. A new
processing ow is adopted in this printing system.
To deal with various fabrics, different algorithms are proposed,
shape context based one for embroidery with only stitches along
shape, random forest model and Gabor features based one for full
embroidery and torque distribution based one for lace. The
experiments demonstrate that the proposed algorithms obtain
great points correspondences and image alignment results. This
prototype can also produce well-designed customized fabrics.
Possible future work includes extending these methods for
more types of fabrics, improving productivity and updating
printing system to automatically run the corresponding algorithm
for fabrics without affecting its accuracy.
Acknowledgement
This work was supported by the National Natural Science
Foundation of China (grant number 61472348).
References
[1] H. Ujiie, Digital Printing of Textiles, Woodhead Publishing, Cambridge, 2006.
[2] D.J. Tyler, Textile digital printing technologies, Text. Prog. 37 (4) (2005) 165.
[3] S. Gupta, Inkjet printing a revolutionary ecofriendly technique for textile
printing, Indian J. Fibre Text. Res. 26 (1/2) (2001) 156161.
[4] K.J. Fang, S.H. Fu, X. Zhang, et al., Development of digital ink jet printing,
Dyeing Finish. 24 (2004) 4851.
[5] I. Petrinic, S.S. Turk, B. Neral, Digital textile printing, Pigment Resin Technol. 50
(7) (2001) 351356.
[6] M.N. Allen, Combination embroidery/screen printing apparatus and method,
Patent 5,144,899, USA, 1992.
[7] J. Buck, System, apparatus, and method for integrating screen printing and
embroidery on a garment while mounted in a single hoop, Patent 6,957,614,
USA, 2005.
[8] M. Kuki, H. Kitazawa, Embroidery sewing machine with printing function,
Patent 7,715,939, USA, 2010.
[9] R. Szeliski, J. Coughlan, Spline-based image registration, Int. J. Comput. Vis. 22
(3) (1997) 199218.
[10] B. Zitova, J. Flusser, Image registration methods: a survey, Image Vis. Comput.
21 (11) (2003) 9771000.
[11] E. Knecht, J.B. Fothergill, The Principles and Practice of Textile Printing, Charles
Grifn, London, 1912.
[12] K. Siemensmeyer, B. Siegel, S. Ervine, et al., Solutions for digital textile printing,
NIP and digital fabrication conference, Soc. Imaging Sci. Technol. (1) (1999)
280283.
[13] J. Weiser, F.W. Raulfs, K. Siemensmeyer, Digital textile printing NIP and digital
fabrication conference, Soc. Imaging Sci. Technol. (2) (2000) 529532.
[14] J. McCann, D. Bryson, Smart Clothes and Wearable Technology, Woodhead
Publishing, Cambridge, 2009.
[15] X. Dai, S. Khorram, A feature-based image registration algorithm using
improved chain-code representation combined with invariant moments, IEEE
Trans. Geosci. Remote Sens. 37 (5) (1999) 23512362.
[16] O. Cordn, S. Damas, J. Santamara, Feature-based image registration by means
of the CHC evolutionary algorithm, Image Vis. Comput. 24 (5) (2006) 525533.
[17] L. Huang, Z. Li, Feature-based image registration using the shape context, Int. J.
Remote Sens. 31 (8) (2010) 21692177.
[18] M. Carfagni, R. Furferi, L. Governi, A real-time machine-vision system for
monitoring the textile raising process, Comput. Ind. 56 (89) (2005) 831842.
[19] W.Y. Wu, M.J.J. Wang, C.M. Liu, Automated inspection of printed circuit boards
through machine vision, Comput. Ind. 28 (2) (1996) 103111.
[20] F. Bianconi, L. Ceccarelli, A. Fernandez, et al., A sequential machine vision
procedure for assessing paper impurities, Comput. Ind. 65 (2) (2014) 325332.
[21] L. Breiman, Random forests, Mach. Learn. 45 (1) (2001) 532.
[22] N. Otsu, A threshold selection method from gray-level histograms, Automatica
11 (1975) 285296.
[23] W. Wang, Y.S. Wong, G.S. Hong, Flank wear measurement by successive image
analysis, Comput. Ind. 56 (8) (2005) 816830.
[24] S. Belongie, J. Malik, J. Puzicha, Shape matching and object recognition using
shape contexts, IEEE Trans. Pattern Anal. Mach. Intell. 24 (4) (2002) 509522.
[25] F.L. Bookstein, Principal warps: thin-plate splines and the decomposition of
deformations, IEEE Trans. Pattern Anal. Mach. Intell. (6) (1989) 567585.
[26] T. Ahonen, A. Hadid, M. Pietikainen, Face description with local binary
patterns: application to face recognition, IEEE Trans. Pattern Anal. Mach. Intell.
28 (12) (2006) 20372041.
[27] O. Ghita, P.F. Whelan, T. Carew, et al., Quality grading of painted slates using
texture analysis, Comput. Ind. 56 (8) (2005) 802815.
[28] A.K. Jain, F. Farrokhnia, Unsupervised texture segmentation using Gabor lters,
Pattern Recognit. 24 (12) (1991) 11671186.
[29] R. Jenssen, T. Eltoft, Independent component analysis for texture
segmentation, Pattern Recognit. 36 (10) (2003) 23012315.
[30] Y. Xu, Y. Quan, Z. Zhang, et al., Contour-based recognition, IEEE Conference on
Computer Vision and Pattern Recognition, (2012) , pp. 34023409.
[31] A. Myronenko, X. Song, Intensity-based image registration by minimizing
residual complexity, IEEE Trans. Med. Imaging 29 (11) (2010) 18821891.
[32] S. Klein, U.A. van der Heide, I.M. Lips, et al., Automatic segmentation of the
prostate in 3D MR images by atlas matching using localized mutual
information, Med. Phys. 35 (4) (2008) 14071417.
[33] R.C. Veltkamp, Shape matching: similarity measures and algorithms, IEEE
International Conference on Shape Modeling and Applications, (2001) , pp.
188197.
[34] A. Jimeno-Morenilla, J. Garcia-Rodriguez, S. Orts, et al., GNG based foot
reconstruction for custom footwear manufacturing, Comput. Ind. 75 (6) (2016)
116126.
[35] J.P. Pernot, B. Falcidieno, F. Giannini, et al., Incorporating free-form features in
aesthetic and engineering product design: state-of-the-art report, Comput.
Ind. 59 (6) (2008) 626637.