Journal of Mechanical Science and Technology 37 (8) 2023 DOI 10.

1007/s12206-023-0740-y

Journal of Mechanical Science and Technology 37 (8) 2023

Original Article
DOI 10.1007/s12206-023-0740-y
An autonomous system design for mold
loading on press brake machines using a
Keywords:
· Deep learning
· Image processing
camera platform, deep learning, and
· Mold
· Press brake
image processing
· YOLOv4
Muhammet Üsame Öziç1, Mücahid Barstuğan2 and Atakan Özdamar3
1
Correspondence to: Department of Biomedical Engineering, Faculty of Technology, Pamukkale University, Denizli, Turkey,
Muhammet Üsame Öziç 2
Department of Electrical-Electronics Engineering, Faculty of Engineering and Natural Sciences, Konya
[email protected] 3
Technical University, Konya, Turkey, Research and Development Center (R&D), MVD Machinery Indus-
try Inc., Konya, Turkey
Citation:
Öziç, M. Ü., Barstuğan, M., Özdamar, A.
(2023). An autonomous system design for Abstract Press brakes are among the most important machines used in sheet metal
mold loading on press brake machines processing. In these machines, different numbers of molds are used for sheet bending and
using a camera platform, deep learning,
and image processing. Journal of Me-
these molds are placed in the system by an operator. However, this process is slow, error-
chanical Science and Technology 37 (8) prone, and dependent on human labor. In this study, a real-time system that automatically de-
(2023) 4239~4247. tects molds and manipulates a robotic arm was designed using YOLOv4 and image process-
http://doi.org/10.1007/s12206-023-0740-y
ing. YOLOv4, a deep learning (DL)-based object detection algorithm, was applied to detect the
positions, types, and holes of molds. Classical image processing methods were implemented to
Received November 6th, 2022 find the center (X, Y) coordinates of the mold hole. This study shows that the press brake ma-
Revised March 27th, 2023 chines currently used in industry can be transformed into smart machines through DL, image
Accepted May 5th, 2023 processing, camera systems, and robotic arm features.

† Recommended by Editor
Hyung Wook Park

1. Introduction
Press brakes are CNC-controlled metal bending machines often used in industrial applica-
tions such as the production of automotive panels, aircraft fuselage, furniture, and others.
Sheet metals are produced as flat shapes and can be transformed by press brake machines
into preferred shapes in a short time. The weight of press brakes vary in the range of 30-20000
tons and the width in the range of 1000-8000 mm. Two types of molds are used in the bending
process: upper and lower. The upper and lower molds have the same “V” shape to allow the
upper mold to enter the lower mold. This design bends the sheet between the molds with high
hydraulic power according to the production schedule [1-4]. Molds used in metal sheet bending
have varying sizes, dimensions, and numbers. An operator places these molds in the machine
according to conventional methods. Then, the operator places the sheet metal, which is bent,
between the molds and implements the desired process using hydraulic power.
With the development of robotic technology in recent years, industrial robotic arms have been
used to feed press brakes. Integrating industrial robotic arms into press brake machines in-
creases production flexibility and speed. Thus, the machines are transformed into successful
robotic automation applications with reduced cycle time and higher quality [5-8]. Worldwide
manufacturers are developing robotic arm-fed press brake machines. The feeding system in-
cludes placing molds and sheet metals into machines according to software. Studies on press
brake machines and robotic systems are based on the loading and unloading of sheets and
bending of the sheet metal with robotic arms [5-8]. Other studies are mostly focused on me-
chanical strength and system modeling [9], determining material properties [10], hydraulic sys-
tems [11], pressure modeling [12], different press brake designs [13-15], optimization of the
© The Korean Society of Mechanical bending process [16], finite element analysis [17], determination of bending parameters [18],
Engineers and Springer-Verlag GmbH
Germany, part of Springer Nature 2023 and fatigue analysis of sheet metals shaped by the press brake [19]. The camera and sensor

4239
 Journal of Mechanical Science and Technology 37 (8) 2023 DOI 10.1007/s12206-023-0740-y

systems are used to measure the bending angle after the

process [20, 21].
In this study, a real-time preliminary application was per-
formed to convert a conventional press brake machine into a
smart machine using deep learning (DL), image processing, a
camera platform, and a robotic arm. The experimental setup
was created by a camera and illumination system on the ro-
botic arm. The autonomous mold loading application program
was implemented by using Python software, OpenCV, the
Skimage image processing library, and the YOLOv4 DL algo-
rithm. Training and testing processes were performed for mold
position, type, and hole detection. Mold center hole coordinates
were determined by conventional image processing methods
after automatic detection of the YOLOv4 algorithm. Then, ac-
cording to these coordinate values, the loading process was
carried out from the molded panel to the gripper of the robotic
arm. This study was performed at a factory of MVD Machinery
Industry (https://www.mvd.com.tr), a manufacturer of press Fig. 1. Bending of sheet metal using mold and pressure.
brake machines in Konya, Turkey. In this study, a new image
processing-based approach has been investigated as an alter-
native to the fixed program mold loading robot designed in the
factory [22].

2. Material and methods

2.1 Mold types
The mold is the most basic component of press brake ma-
chines. Nevertheless, it is considered a small accessory during
sheet metal bending. Even if the press brake machines be-
come highly sensitive, the main task in sheet bending is done
by mold types. The molds are placed on the press brake ma-
chines with various systems such as mechanical, pneumatic,
Fig. 2. The platform on which the molds with different numbers are placed.
and hydraulic grippers. The molds that have different geome-
tries make the shaping of the sheet metal easier by bending
and manufacturing according to the industrial standard lengths arrayed molds, a robotic arm, a press brake machine, a cam-
accepted. The molds, which enable bending at different angles, era, an illumination system, and a computer. The molds were
can perform the bending of complex parts within the specified arranged as shown in Fig. 2. As the environmental light was
tolerance range. The molds are produced in standard lengths constantly changing in the factory, a ring LED-shaped illumina-
accepted in the industry. Certain depths in the upper parts are tion panel was used to fix the light. Environmental lights were
set according to the bending process needed. The molds are suppressed in the setup where only the front illumination was
generally manufactured from C45 fabrication steel; however, used. A USB camera was placed in the middle of the ring illu-
some molds are manufactured from 42CrMo4 (4140) tempered mination so that the camera could take the reflected light ho-
steel. The surfaces of the molds in contact with the part to be mogeneously. The illumination system and camera were
bent are subjected to heat treatment and hardening. Hardened mounted on a six-axis Mitsubishi Electric RV-20FRM robotic
molds have a longer life. Fig. 1 shows the bending of sheet arm. During the system operation, because the reach distance
metal by applying pressure with a mold. Fig. 2 shows the of the robotic arm to the press brake, the molded panel, and
molds numbered 20, 25, 30, 35, 40, and 100 used in the press the separation area of the bent sheet metals was not enough,
brake manufacturing factory where the study was carried out. the robotic arm was placed on a linear slide. This sliding linear
DL and image processing applications were performed on the mechanism provided a robotic arm to reach all production ar-
panel with molds presented in Fig. 2. eas in the system. The pneumatic gripper was mounted on the
robotic arm to load the molds. The camera and gripper were
centered on the Y-axis. Thus, the gripper, camera, and robotic
2.2 Experimental setup
arm configurations were performed at a highly precise rate.
The robotic arm-fed-based experimental setup consisted of After the experimental setup, robotic arm movement and image

4240
 Journal of Mechanical Science and Technology 37 (8) 2023 DOI 10.1007/s12206-023-0740-y

the annotation for the mold position, Fig. 4(b) shows the anno-
tation for the mold type, and Fig. 4(c) shows the annotation for
holes on the various molds.

2.4 YOLOv4
DL models are powerful algorithms that are frequently used
in object detection, segmentation, and recognition [23]. The
development of DL methods, which started with the recom-
mendation of LeNet in 1998, stopped for a period because of
hardware inadequacies and continued with the introduction of
AlexNET, VGG, GoogleNET, and ResNET in 2012 [24-28].
Fig. 3. Experimental setup. After these developments, which are generally used for image
classification, the need to detect the object on an image quickly
without any size difference has emerged. With the recommen-
dation of SSD, R-CNN, Fast R-CNN, and Faster R-CNN mod-
els based on the region-based CNN approach, fast algorithms
with high detection rates have entered the Refs. [29-33]. These
methods are models that search for the most suitable object
with high accuracy by dividing the image into regions. Another
region-based DL model is the YOLO algorithm [34]. YOLO, a
CNN-based DL model, is a fast enough algorithm to detect
(a) (b)
objects on the image at one time. YOLO is constantly improv-
ing itself with different versions. In this study, the YOLOv4 algo-
rithm, which entered the literature in April 2020, was used. As
the processing load is very high in classical CNN methods,
YOLOv4 has been successful compared to its competitors in
real-time applications [35]. The YOLOv4 method uses regres-
sion to describe object positioning with bounding boxes. This
algorithm first divides the image into regions and searches for
(c)
the desired object within a specific region. The confidence
Fig. 4. (a) Image annotation for mold position detection; (b) image annota- score is determined by calculating the probability of the
tion for mold type detection; (c) image annotation for mold hole detection. searched object being in that region. This score gives the per-
centage of probability that the searched object is in that area.
processing systems operated collaboratively. MODBUS TCP/ However, overdetection and drawing several bounding boxes
IP protocol communicated with the robotic arm and Python around an object can occur in experiments. To prevent this
software. Fig. 3 shows the experimental setup of the robotic problem, a non-maximum suppression technique is applied to
arm-fed press brake machine. the objects inside the bounding boxes. This technique excludes
objects with a lower confidence score from evaluation and
2.3 Dataset and image annotation checks for the presence of a higher confidence bounding box
in the same region [35, 36]. YOLOv4 runs on the DarkNET
In this study, three different datasets were prepared for de- framework, which was developed with C/CUDA and has an
tection of mold positions, types, and holes. For mold position excellent performance. Therefore, the DarkNET framework
and mold type detection, a two-minute video was taken from a (https://github.com/AlexeyAB/darknet) was first downloaded to
distance of 40 cm using the USB camera on the robotic arm. the computer and the installation process was completed.
These videos were recorded frame by frame with a Python
script, and then two image datasets were created in .jpg format.
2.5 Network settings
To detect the mold hole, the robot arm was brought closer to
the molds and a two-minute video was taken and recorded as All three datasets were created from 1250 images in .jpg for-
images. The reason for this process is that after mold position mat. The data were divided into 75 % training and 25 % testing
and mold number detection, the robot moves closer and de- sets. Images with .txt files containing object coordinate values
tects the mold holes. Each dataset was annotated using the were saved in three different DarkNET folders. As YOLOv4 is
free browser-based “makesense” application (https://www. currently trained with the COCO dataset with 80 outputs, it
makesense.ai/). The annotated data were saved in YOLO for- should be customized according to the number of classes to be
mat as .txt files containing coordinate values. Fig. 4(a) shows used in this study. The pre-trained weights were downloaded

4241
 Journal of Mechanical Science and Technology 37 (8) 2023 DOI 10.1007/s12206-023-0740-y

as “yolov4.conv.137” and saved in the DarkNET folder average of the sum of the AP values of all classes (Eq. (6)).
(https://github.com/AlexeyAB/darknet). Some network hyper- The higher the AP and mAP, the better the detection perform-
parameters in the “yolov4.cfg” configuration file in the DarkNET ance of the model [37, 38].
folder have been changed for transfer learning according to the
purpose of the training. The number of filters according to the 1

APk = ∫ Pk ( Rk )dRk (5)

number of classes was calculated with the formula “filters = 0
3*(c+5)”. The number of classes has been changed to one for 1 k
mold position detection (1 class), one for mold hole detection (1 mAP = ∑ APi .
k i=1
(6)
class), and six for mold type detection (6 classes). Several
parameters in the “Makefile” file within the DarkNET framework
have been changed (GPU = 1, CUDNN = 1, OPENCV = 1) 2.7 Image processing
because the Google virtual computer was used over COLAB.
Image processing processes were used in two stages in this
After these processes were completed, the .zip file was up-
study. In the first stage, the weights obtained during the training
loaded to Google Drive and prepared for training and testing.
process and the configuration file used were given as input to
the “cv2.dnn” method of the OpenCV library. This method is a
2.6 Performance evaluation function that comes with the OpenCV 4.0 version for the practi-
cal use of DL algorithm weights. The coordinates of the tar-
Network performance evaluations were performed with pre-
geted object on the image were found with this method and
cision (P), recall (R), F1-score (F1-S), intersection of union
bounding boxes were drawn. The “cv2.dnn.NMSBoxes” meth-
(IoU), and mean average precision (mAP). The IoU value is a
od was used to suppress the boxes with a confidence score
metric that indicates the extent of overlap between the bound-
below a certain threshold with the non-maximum suppression
ing box annotated object (A) and the detected object (B). The
method. The box with the highest confidence score was drawn
IoU value is calculated as the area where the two rectangles
on the image and printed. These methods were applied using
intersect, divided by the area of the union of these two rectan-
three separate Python scripts for mold position, mold type, and
gles (Eq. (1)). An IoU value above 0.5 is considered to have a
mold hole detection. Finding the hole center point exactly after
successful detection process. Otherwise, the detection process
the mold hole detection process constitutes the second stage
fails.
of image processing. The biggest problem here was the notch
that protrudes under the hole. The region of interest obtained
A∩ B by cropping the relevant region determined by hole detection
IoU = . (1)
A∪ B was passed through the OTSU threshold, hole filling, morpho-
logical processes, and central coordinate finding processes,
The estimated number of bounding boxes above the IoU 0.5 respectively.
value is calculated as true positive (TP), and the presence of
undesired bounding boxes is calculated as false positive (FP).
False negative (FN) is the number of regions that the system 3. Results
could not find even though an annotated region existed. True 3.1 Training and testing results
negative (TN) is not considered in object detection perform-
ance calculations. The precision (P) calculation specifies the In Table 1, the results of performance criteria for mold posi-
ratio of TP detections to all detections (Eq. (2)). Recall (R) tion detection, mold hole detection, and mold type detection
value is determined as the ratio of TP detected bounding boxes are given. Training of each dataset took about 25 hours and
to all bounding boxes that need to be detected (Eq. (3)). It is was completed in 4000 iterations. The P, R, F1-S, and mAP
the harmonic average of the F1 score (F1-S), precision, and values were obtained as 100 % for each detection process.
recall values given in Eq. (4). These values showed that the test process had been success-
fully achieved with high accuracy. The IoU values were calcu-
lated as 92.78 % for mold position, 92.12 % for mold hole, and
TP
Precision = (2) 88.57 % for mold number. The IoU value above 50 % indicated
TP + FP
that the detection was successful. Therefore, the bounding box
TP
Recall = (3)
TP + FN
Precision * Recall Table 1. Results of performance criteria for training and testing.
F1 − score = 2 * . (4)
Precision + Recall
Class P R F1-S IoU mAP
Mold position 1 100 % 100 % 100 % 92.78 % 100 %
To evaluate the performance of the model in detecting each
Mold hole 1 100 % 100 % 100 % 92.12 % 100 %
class (k), the area under the precisionk-recallk graph is defined
Mold number 6 100 % 100 % 100 % 88.57 % 99.98 %
as the average precision (APk) (Eq. (5)). Mean AP (mAP) is the

4242
 Journal of Mechanical Science and Technology 37 (8) 2023 DOI 10.1007/s12206-023-0740-y

(a) (b)

(c) (d)

(e) (f)

Fig. 5. (a) Mold position detection without NMS; (b) mold position detection using NMS; (c) mold number detection without NMS; (d) mold number detection
using NMS; (e) mold hole detection without NMS; (f) mold hole detection using NMS.

area in the labeling performed by the operator and the bound- method of the OpenCV library was used for this process.
ing box area detected by YOLOv4 largely overlapped. The Fig. 5(a) shows the mold position detection without NMS,
limitation of the training process was that the Google COLAB and Fig. 5(b) shows the mold position detection using NMS
environment interrupted the user from the server after a certain and confidence score. Fig. 5(c) shows the mold type detection
working period. Therefore, the code scripts were run again to without NMS, and Fig. 5(d) shows the mold type detection
continue the training process from where it left off. However, using NMS and confidence scores. Fig. 5(e) shows the mold
because of these interruptions, the loss and mAP graphs of the hole detection without NMS, and Fig. 5(f) shows the mold hole
training process could not be drawn. detection using NMS and confidence score. As the last task,
the gripper must enter the middle of the mold hole. However,
the midpoint of the mold hole detected by YOLOv4 does not
3.2 Visual results
represent the hole center coordinates. The center coordinates
After obtaining the YOLOv4 weights, automatic object detec- were determined by traditional image processing methods.
tion experiments were conducted on the test data using the After the mold type and position detection, the robotic arm was
“cv2.dnn” module. Bounding boxes were drawn according to brought closer to the relevant mold. The mold hole was deter-
the estimated object coordinates. However, the algorithm can mined and bounding box coordinate values were recorded.
find the bounding box surrounding more than one object in the The color image was converted to a gray level and cropped
same region in the image divided into the SxS grid. The non- from the bounding coordinate values obtained in the previous
maximum suppression method was used to obtain the one with step. The image was converted to binary format by finding the
the highest confidence score from more than one bounding appropriate threshold with the Otsu thresholding method. Black
box representing the same region. The “cv2.dnn.NMSBoxes” holes in the object were filled. The notch under the gap pre-

4243
 Journal of Mechanical Science and Technology 37 (8) 2023 DOI 10.1007/s12206-023-0740-y

Fig. 6. Image processing flowchart to determine center coordinates.

Fig. 7. Center coordinates (red points) of different molds obtained as a result of image processing and YOLOv4.

vents the exact middle point from being found. For this reason, hardware design, calibration, and communication protocols of
the morphological opening process was carried out using a the products were completed before the system was started.
disk structure element of r = 240 size. The center coordinates The entire system can be controlled through the main Python
of the obtaining round binary object were practically found with script developed. In the first step, the operator enters the de-
OpenCV’s contour and moment operations. In Fig. 6, the flow sired mold number into the system. In the second stage, the
diagram of the classical image processing methods performed previously trained mold position and mold number detection
after the mold hole detection process is given. In Fig. 7, the weights are automatically executed to obtain the information of
hole center coordinate values and drawings obtained in differ- all molds. Position information according to the mold number
ent molds are given. entered by the operator is sent to the robot using MODBUS
TCP/IP communication protocol and the robotic arm gets clos-
er to this mold. The mold hole region is detected automatically
3.3 Autonomous robotic manipulation
using YOLOv4 mold hole detection weights. According to the
In this section, the flow diagram of the DL and image proc- pipeline given in Fig. 6, classical image processing methods
essing-based system for the robotic arm to take the mold is are applied to this region, and center coordinate values are
explained. The system consists of two parts: Python + Image obtained. These values are sent to the robotic arm and the
Processing + YOLOv4 and Mitsubishi robotic software. The mold is entered according to the center coordinate values of

4244
 Journal of Mechanical Science and Technology 37 (8) 2023 DOI 10.1007/s12206-023-0740-y

contraction depending on the ambient temperature. Therefore,

gripper breaks occur because of coordinate shifts in manually
programmed robots. Relying on an operator to place the molds
is a time-consuming and error-prone task. Thus, in this study,
an application has been carried out to take the molds with ro-
bots powered by DL and image processing methods. As the
proposed system updates the mold coordinates in every cycle,
errors caused by the operator and environmental variables are
eliminated. In the study, in which the YOLOv4 DL algorithm
was used, high accuracy rates were obtained with mold posi-
tion detection 100 % mAP and 92.78 IoU, mold number detec-
tion 99.98 % mAP and 88.57 % IoU, mold hole detection
100 % mAP and 92.12 % IoU performance values.
With the coordinate values determined by YOLOv4 and im-
age processing, the robot arm can enter the mold center and
perform the loading process. As a result of 50 separate ex-
periments, only one error was recorded through camera com-
munication. Thus, the system’s success was achieved at a
high accuracy rate of 98 %. Although the designed system
provides remarkable results for mold detection experiments,
problems remain in machine integration. The image on the
upper right corner of Fig. 2 shows that the numbers of the two
frequently used molds have been deleted. Therefore, because
these two mold numbers cannot be determined, they cannot be
taken with the robotic arm. Furthermore, some molds have
traces of dirt and oil. These problems hinder DL, image proc-
essing, and robotic manipulation. For robotic manipulation to
work in ideal conditions, the mold numbers have to be clearly
visible and free from external elements such as oil and dirt.
Another limitation is the use of appropriate hardware to in-
crease detection speed. As speed is an important factor in
industrial processes, detection time should be minimized.
Fig. 8. Flowchart of robotic manipulation. Future studies will be conducted to increase the speed by
using NVIDIA Jetson cards with powerful GPUs. In the litera-
the gripper and the mold is loaded from the panel. In Fig. 8, the ture, studies in which camera and sensor systems are used in
flow diagram of the robotic manipulation process is shown. press brakes have reported the calculation of the bending an-
This flowchart describes a cycle of autonomous robotic ma- gle after the bending process [20, 21]. To the best of our knowl-
nipulation. For the system to work again and load a different edge, this study is the first to describe an AI-based design that
mold, the operator must repeat the data entry. automatically loads molds on press brake machines using a
camera system, a robotic arm, DL, and image processing.
4. Discussion
With Industry 4.0, the development of image processing and
5. Conclusion
machine learning-based robotic systems has begun to ad- In this study, a real-time industrial application based on DL
vance in industrial applications. In studies where classical im- and image processing has been designed to autonomously
age processing methods are used, conditions such as varying take molds in press brake machines with a robotic arm. The
environmental light, an excessive number of patterns, and integration of artificial intelligence and camera systems into
different objects in the area scanned by the camera cause industrial applications is becoming more common, minimizing
difficulty in detecting the targeted industrial object. Through the the problems caused by human error and offering fast and
use of powerful DL algorithms, practical solutions can be practical solutions. This study shows how a conventional press
achieved in difficult tasks such as object detection, object clas- brake machine can be transformed into an intelligent system
sification, and object recognition in industrial applications. In with DL and image processing methods using a camera sys-
press brakes, mold feeding systems are programmed manually tem. In future studies, system optimization, full automatic bend-
or molds are placed in the system by a human operator. Mold ing applications without the need for human intervention, and
coordinates can change as a result of panel expansion and automatic determination of the bending angle in the sheet

4245
 Journal of Mechanical Science and Technology 37 (8) 2023 DOI 10.1007/s12206-023-0740-y

metal with image processing will be investigated. brake, Journal of Materials Processing Technology, 141 (1)
(2003) 143-154.
[11] Y. Altintaş and A. J. Lane, Design of an electro-hydraulic CNC
Acknowledgment press brake, International Journal of Machine Tools and Manu-
We would like to thank Mr. A. Vasıf Inan, the factory owner, facture, 37 (1) (1997) 45-59.
for allowing this study to be conducted at MVD Machinery In- [12] P. L. Teixeira, W. Vianna Jr, R. D. Penteado, P. Krus and V. J.
dustry. De Negri, Pressure modeling and analysis of a synchronized
hydraulic press brake with variable-speed pump, Proceedings
of the ASME/BATH 2015 Symposium on Fluid Power and Mo-
Nomenclature------------------------------------------------------------------ tion Control. ASME/BATH 2015 Symposium on Fluid Power
AP : Average precision and Motion Control, Chicago, USA (2015) V001T01A067.
F1-S : F1-score [13] N. Gwangwava, K. Mpofu, N. Tlale and Y. Yu, Sheet metal
IoU : Intersection of union productivity improvement through a new press brake design,
mAP : Mean average precision African Journal of Science, Technology, Innovation and Devel-
NMS : Non-maximum suppression opment, 6 (2) (2014) 135-144.
P : Precision [14] D. Mourtzis, V. Zogopoulos, I. Katagis and P. Lagios, Aug-
R : Recall mented reality based visualization of CAM instructions towards
YOLO : You only look once Industry 4.0 paradigm: a CNC bending machine case study,
DL : Deep learning Procedia CIRP, 70 (2018) 368-373.
TP : True positive [15] E. Nikolidakis and G. C. Vosniakos, A novel tool clamping
FP : False positive system for CNC press brakes, Proceedings in Manufacturing
TN : True negative Systems, 12 (2) (2017) 59.
FN : False negative [16] N. Kontolatis and G.-C. Vosniakos, Optimisation of press-
brake bending operations in 3D space, Journal of Intelligent
Manufacturing, 23 (3) (2012) 457.
References
[17] J. Bessa Pacheco and A. D. Santos, A study on the nose
[1] T. Altan and A. E. Tekkaya, Sheet Metal Forming: Fundamen- radius influence in press brake bending operations by finite
tals, 1st Ed., ASM International, Ohio, America (2012). element analysis, Key Engineering Materials, 554 (2013) 1432-
[2] R. Fournier, Sheet Metal Handbook: How to Form and Shape 1442.
Sheet Metal for Competition, Custom and Restoration Use, HP [18] F. Aydemir, B. G. Cakan, A. Durmuş and K. Çavdar, Determi-
Books, 575 (1989). nation of press brake bending parameters for hardox 400 steel,
[3] B. L. Rapien and S. Fournier, Fundamentals of Press Brake Uludağ University Journal of The Faculty of Engineering, 27 (1)
Tooling, 2nd Ed., München:Hanser Publishers (2010). (2022) 325-340.
[4] K. A. Stelson, An adaptive press brake control for strain- [19] K. E. Barth, G. K. Michaelson and R. M. Tennant, Fatigue
hardening materials, Journal of Engineering for Industry, 108 performance of singular and modular press-brake-formed steel
(2) (1986) 127-132. tub girders, Bridge Structures, 16 (1) (2020) 3-13.
[5] X. Yin, H. Wang and G. Wu, Path planning algorithm for bend- [20] R. Yan, L. Zhang, Z. Zhao, Y. Zhang and Y. Wang, A dynamic
ing robots, IEEE International Conference on Robotics and sheet metal bending angle measurement method based on
Biomimetics (ROBIO), IEEE, Guilin, China (2009) 392-395. line structured lasers, Lasers in Engineering, 38 (2017) 93-107.
[6] F. Xu, Q. Jiang, L. Rong, P. Zhou and J. Hu, Structural model [21] S. Raftopoulos and G. C. Vosniakos, A cost-effective device
and dynamic analysis of six-axis cartesian coordinate robot for for online measurement of sheet metal bending angle on CNC
sheet metal bending, International Journal of Advanced Ro- press brakes based on capacitive sensors, International Jour-
botic Systems, 16 (4) (2019). nal of Mechatronics and Manufacturing Systems, 13 (2) (2020)
[7] N. Kontolatis, G. C. Vosniakos and K. Kyriakopoulos, On pa- 153-168.
rametric toolpath design of a robot serving a press-brake, Pro- [22] MVD Machinery Industry Inc., Press Brake With Robot, YouTube,
ceedings 19th Flexible Automation and Intelligent Manufactur- https://www.youtube.com/watch?v=jna2zEnJUhk&ab_channel=
ing Conference, Middlesbrough, United Kingdom (2009) 6-8. MVD.
[8] S. K. Gupta, D. A. Bourne, K. H. Kim and S. S. Krishnan, Auto- [23] M. Hassaballah and A. I. Awad, Deep Learning in Computer
mated process planning for sheet metal bending operations, Vision: Principles and Applications, CRC Press (2020).
Journal of Manufacturing Systems, 17 (5) (1998) 338-360. [24] Y. LeCun, L. Bottou, Y. Bengio and P. Haffner, Gradient-
[9] D. Lovrec and M. Kastrevc, Modelling and simulating a con- based learning applied to document recognition, Proceedings
trolled press-brake supply system, International Journal of of the IEEE, 86 (11) (1998) 2278-2324.
Simulation Modelling, 10 (3) (2011) 133-144. [25] A. Krizhevsky, I. Sutskever and G. E. Hinton, Imagenet classi-
[10] R. J. Mentink, D. Lutters, A. H. Streppel and H. J. J. Kals, fication with deep convolutional neural networks, Communica-
Determining material properties of sheet metal on a press tions of the ACM, 60 (6) (2017) 84-90.

4246
 Journal of Mechanical Science and Technology 37 (8) 2023 DOI 10.1007/s12206-023-0740-y

[26] K. Simonyan and A. Zisserman, Very deep convolutional multaneously in orchard using YOLOv4 for robotic pollination,
networks for large-scale image recognition, arXiv:1409.1556 Computers and Electronics in Agriculture, 193 (2022) 106641.
(2014). [38] A. M. Roy, R. Bose and J. Bhaduri, A fast accurate fine-grain
[27] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Angue- object detection model based on YOLOv4 deep neural network,
lov, D. Erhan, V. Vanhoucke and A. Rabinovich, Going deeper Neural Computing and Applications, 34 (5) (2022) 3895-3921.
with convolutions, Proceedings of The IEEE Conference On
Computer Vision And Pattern Recognition, Massachusetts,
USA (2015) 1-9.
[28] K. He, X. Zhang, S. Ren and J. Sun, Deep residual learning Muhammet Üsame Öziç graduated
for image recognition, Proceedings of The IEEE Conference from the Department of Electronics Engi-
On Computer Vision And Pattern Recognition, Las Vegas, neering of Uludağ University with a B.Sc.
USA (2016) 770-778. degree in 2010, from the Department of
[29] P. Bharati and A. Pramanik, Deep learning techniques—R- Electrical and Electronics Engineering of
CNN to mask R-CNN: A survey, Computational Intelligence in Selcuk University with an M.Sc. degree
Pattern Recognition: Proceedings of CIPR (2020) 657-668. in 2013, and the Electrical & Electronics
[30] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C. Y. Fu Engineering Department of Selcuk Uni-
and A. C. Berg, SSD: single shot multibox detector, Computer versity with a Ph.D. degree in 2018. His professional interests
Vision - ECCV 2016. ECCV 2016. Lecture Notes in Computer include deep learning, MRI-based frameworks, medical image
Science, Springer, Cham, 9905 (2016) https://doi.org/10.1007/ analysis, image processing, image reconstruction, robotics,
978-3-319-46448-0_2. and pattern recognition.
[31] R. Girshick, Fast R-CNN, Proceedings of The IEEE Interna-
tional Conference on Computer Vision, Santiago, Chile (2015) Mücahid Barstuğan graduated from the
1440-1448. Department of Electronics Communica-
[32] S. Ren, K. He, R. Girshick and J. Sun, Faster R-CNN: to- tion Engineering at Kocaeli University
wards real-time object detection with region proposal networks, with B.Sc. degree in 2010, from the Elec-
IEEE Transactions on Pattern Analysis And Machine Intelli- trical & Electronics Engineering Depart-
gence, 39 (6) (2017) 1137-1149. ment of Selcuk University with an M.Sc.
[33] R. Girshick, J. Donahue, T. Darrell and J. Malik, Rich feature degree in 2014, and Electrical & Electron-
hierarchies for accurate object detection and semantic segmen- ics Engineering Department of Konya
tation, Proceedings of The IEEE Conference on Computer Vi- Technical University with a Ph.D. degree in 2019. His profes-
sion And Pattern Recognition, Ohio, USA (2014) 580-587. sion includes machine learning, Mitsubishi robotic systems and
[34] J. Redmon, S. Divvala, R. Girshick and A. Farhadi, You only programming, image processing, image segmentation, image
look once: unified, real-time object detection, Proceedings of reconstruction, and PLC programming.
The IEEE Conference on Computer Vision And Pattern Rec-
ognition, Las Vegas, USA (2016) 779-788. Atakan Özdamar graduated from the
[35] A. Bochkovskiy, C. Y. Wang and H. Y. M. Liao, YOLOV4: Department of Electrical & Electronics
optimal speed and accuracy of object detection, arXiv:2004. Engineering with a B.Sc. degree in 2019.
10934 (2020). He is a research engineer at MVD Ma-
[36] D. Wang, C. Li, S. Wen, Q. L. Han, S. Nepal, X. Zhang and Y. chinery Industry, Research and Devel-
Xiang, Daedalus: breaking nonmaximum suppression in object opment Center (R&D) in Konya, Turkey.
detection via adversarial examples, IEEE Transactions on Cy- At the same time, he is pursuing his the-
bernetics, 52 (8) (2021) 7427-7440. sis work as a graduate student at the
[37] G. Li, R. Suo, G. Zhao, C. Gao, L. Fu, F. Shi, J. Dhupia, R. Li Department of Electrical and Electronics Engineering at Konya
and Y. Cui, Real-time detection of kiwifruit flower and bud si- Technical University.

4247