Professional Documents
Culture Documents
iew
Classification and Localization of Surface Defects
Manjeet Kaur1 , Krishan Kumar Chauhan1 , Isibor Kennedy Ihianle2 ,
Kayode Owa2 , Naveen Aggarwal1 , Renu Vig1 , Garima Joshi1∗
1
UIET, Panjab University, Chandigarh, India
ev
2
School of Science and Technology, Nottingham Trent University,
50 Shakespear Street, Nottingham, NG1 4FQ, United Kingdom
r
Abstract
er
Quality inspection of fasteners is compulsory for safe and reliable operations.
Manual inspection is usually necessary to assure product quality, but this
results in production bottlenecks, lower productivity, and poor efficiency.
pe
To meet the increased demand for high-quality products, sophisticated vi-
sual inspection systems must be integrated into production lines to address
these obstacles. It has the potential to have an influence on industrial sus-
tainability, since it will aid in enhancing industry throughput by identifying
flaws and pinpointing the root cause. In image classification and localization
ot
∗
Corresponding author
Email address: joshi_garima5@yahoo.com ()
Pr
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
1. Introduction
iew
For sustainable operations in fastener manufacturing industry, defects in
fasteners must be recognised and associated with the source of the problems
so that the underlying cause can be determined and corrected [1]. The var-
5 ious types of surface defects caused by mechanical processing, are classified
as crack, wrinkle, dent, scratches, cut, missing and misaligned threads [2]. A
crack is a visible damage along a fastener’s border that is typically caused
ev
by a stressed surface during the heating process [3]. Fastener raw materials
might potentially develop cracks. Wrinkles are a form of imperfection on
10 fasteners caused by material displacement during the forging process of nuts
in particular [4]. Another kind of defect is a dent, which happens when metal
r
is under filled during the forging process, resulting in a deep mark in the sur-
face. Scratches on fasteners is another common defect caused by excessive
15 er
thermal stress on the fastener’s surface during the heat treatment [5]. Cut
on the surface is frequently produced by improper handling of work piece
during tool movement. Damaged threads are also caused by mismanagement
pe
of production tools, resulting in overfilling, misalignment, and mismatched
threads on the fasteners [6]. Thus, an automatic segregation mechanism is
essential to detect these defect.
20 Although image processing and machine learning-based approaches are ef-
ficient but there are certain limitations of these techniques. Images must
be pre-processed, filtered, and scaled applying image processing procedures,
ot
cope with flexibility for noisy and low-quality and resolution images. To
overcome the limitations of all of these strategies, deep learning has been
proven to be much more efficient [8].
The defect detection model proposed in this paper aims to classify as well as
rin
30 localize all the defects present on the surface. From object detection problem
point of view, surface detection needs a trade-off between speed and accuracy
of the model. In the proposed work, an accurate and precise defect classifi-
cation and localization model is targeted rather than a fast but less accurate
ep
model. This paper makes three main contributions: (1) It presents a deep
35 machine learning model for image classification and localization; (2) It evalu-
ates the model on two different datasets containing fasteners of different sizes
and defects; (3) Further, it shows that the application of anchor values for
Pr
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
different aspect ratios significantly improves the classification performance of
iew
the proposed surface defect detection network, with respect to the different
40 features presented. The resulting approach is thus able to be implemented in
real-world automotive user interfaces to provide greater quality inspection of
fasteners. The paper is organised as follows. Section 2 provides a discussion
of related work conducted in the literature. Section 3 describes the proposed
methodology of the study, including a description of the dataset. Section
ev
45 4 presents the experimental results and Section 5 provides the concluding
remarks of the study, as well as proposed future work.
2. Related Works
r
Related work carried out for automatic defect detection system on fas-
teners using deep learning techniques are summarized in this section. Kou
50
er
et al. designed a defect detection model based on YOLOv3 to extract rich
feature information. Dense convolution blocks were included which substan-
tially improved feature characterization ability. On the GC10-DET dataset,
pe
the authors suggested model produced 71.3% mAP, while on the NEU-DET
dataset, it produced 72.2% mAP [9]. Xu et al. integrated the Darknet-53
55 model’s layers with a deep feature neural network. This model was based on
the YOLOv3 model and gives out more than 75% accuracy [10]. In another
study, Gai et al. developed a VGGNet architecture to detect defects on steel
ot
industrial camera. The model achieved 98% accuracy [12]. Patar et al. devel-
oped a faster region-based CNN (faster-RCNN), which is based on the same
screw classes as proposed by Song et al. and is implemented on a Raspberry
65 Pi 3 with a camera module for image capture. The model outperformed the
rin
usual template matching and single shot detection models, with an accuracy
of over 98% [13]. Chen et al. proposed a novel three stage DCNN-based
detection setup, two detectors to localize the cantilever joints and their fas-
teners and a classifier to inspect the fasteners. The proposed model achieved
ep
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
A variety of deep learning-based techniques for defect identification on fas-
iew
75 teners have been investigated after a comprehensive study of the literature
[16]. Both transfer learning-based pre-trained models and developed models
based on pre-trained algorithms have been extensively employed for defect
assessment using deep learning-based techniques [17]. He et al. used 3 aspect
ratios [1:1, 1:2, 2:1] in Region Proposal Network (RPN) which employed to
80 extract region by sliding the feature map, which takes image as an input and
ev
returns a set of rectangular box with an object score as an output [18]. Chen
et al. used k-means clustering to determine the size of anchor boxes in ex-
tended feature pyramid network(EFPN). They used three anchor boxes per
level. Feature maps from the upper layer used a larger anchor box to capture
r
85 a large object, and the lower layers used a smaller anchor box to extract small
objects [19]. The authors observed that YOLOv2 and YOLOv3 model could
not distinguish the small size defects in NEU-DET dataset. They proposed
90
er
end to end defect detection network (EEDN) based on Faster-RCNN which
overcame this problem by using anchor boxes [15].
Although the related works mentioned above are commendable not all de-
pe
fect type, sizes and shapes were considered. A major limitation of YOLOv3
object detection models is its inability to recognise extremely small defects
in higher resolution. The failure to implement models that can match the
industrial quality standards. Furthermore, no machine vision based image
95 dataset for fasteners has been generated so far. Also, single shot detector
ot
3. Proposed Methodology
rin
the resulting images. The images processed, annotated, labelled and then
used as an input of a Surface Defect Detection Network (SSDNet) classifier
with anchor box tuning to optimise performance.
Pr
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
iew
r ev
er
pe
ot
Figure 1: Proposed methodology for machine vision based automatic defect localization
tn
does not have a standard database that is available for this purpose. Hence
for the object detection proposed in this paper, a database specifically for
115 small fasteners (nuts) was created - Small Nut Surface Defect Dataset. The
proposed model has also been tested and validated using a publicly available
ep
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
3.1.1. Small Nut Surface Defect Dataset
Samples of defective and non-defective nuts of size M6 1 (shape: hexago-
iew
120
nal; width across flats: 10mm; thickness: 5.2mm) and M8 (shape: hexagonal;
width across flats: 13mm; thickness: 6.8mm) have been collected from sev-
eral industries for the proposed study work. Five megapixel, Sony IMX264
CMOS sensor machine vision camera with global shutter from LUCID vision
125 under diffused light has been used to gather the samples. Data augmentation
ev
in terms of scaling, contrast variation, rotation resulted in total 1220 images.
Each image has been cropped to the required portion of image to reduce the
complexity of model. Minor details of defects are visible in the images that
have been captured with machine vision camera, a clear difference can be
r
130 seen in Figure 2. The level of complexity in the defect detection is depicted
in Figure 3, varying from the large defects which are clearly visible and the
defects with minor cracks which are difficult to detect. Also, it can be seen
er
that the shape and size of even a similar defect is different. To identify the
pe
ot
tn
Figure 2: a) Image of smallest target Size of Nut b) Microscopic Crack visible with Machine
Vision Camera c) Machine Vision Camera Setup
defect, each image has been annotated and labelled as part of the data pro-
rin
135 cessing in readiness for the classifier. In this study, four classes are targeted:
dent, crack, scratch and non-defective fasteners. To perform labeling of im-
ages, LabelImg tool has been used. A bounding box is created around the
defective surface, which is used to identify and classify the defect. For each
labelled image, an xml file is created, which contains the exact dimension
ep
140 and class label of all the bounding boxes that describe the defect position.
1
Engineers Edge
Pr
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
iew
Figure 3: Level of Complexity in the Dataset
ev
3.1.2. NEU-DET Surface Defect Database
An experiment has been done on a dataset created by Northeastern Uni-
r
versity (NEU) also known as NEU-DET Dataset [15]. The dataset consist of
six types of surface defects of the hot-rolled steel strip. The defect classes in
this dataset involve patches, pitted surface, scratches, rolled-in scale, crazing,
145
er
and inclusion. The dataset consists of 1800 grayscale images, 300 images per
class, resolution of each image is 200x200.
pe
3.2. Surface Defect Detection Network (SDDNet)
A hybrid model called surface defect detection network (SDDNet) is based
150 on single shot detector (SSD) architecture, which takes only one shot to de-
tect all the objects present in an image. Conventional SSD uses MobileNetV2
as the backbone network [20]. MobileNet is a lightweight network that sub-
ot
stantially reduces the network’s complexity, cost and size. It is intended for
real-time object detection. SSD utilizes multi-scale features calculated by
155 feature pyramid network (FPN) layer and default anchor boxes [21]. In the
proposed method, FPN architecture has been used in the Pyramid Structure
tn
165 the final output. SSD head contains 6 prediction layers which are basically
the auxiliary convolutional layers, known as detection head [22]. In detection
head, feature maps are used for predictions. Each box of detection head con-
sists of two parameters: confidence scores, and a box offset/box prediction.
Pr
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
Confidence score tells about the object score in an image, if the object score
iew
170 is higher than the background score, that box will be selected as positive
default box. Each box will be having certain location, size, shape and aspect
ratio, box offset/box prediction adjust all these parameters and generate 4
offset related to original default bounding box [23]. In the proposed model,
the default anchor boxes have been replaced by the new anchor boxes.
r ev
er
pe
Anchor box tuning is one of the important factor that can be adjusted
to boost performance, and optimize the model in terms of accuracy and ef-
ficiency [24]. If anchor boxes are not adjusted as per the need of dataset,
CNN may not be able to detect some small, large, or irregular objects. There-
rin
180 fore, considerable effort can be made to ensure that the correct anchor boxes
are chosen by tuning them according to the need of dataset. The process
described here is used by SSD to tune anchors. First step is to create multi-
ple anchor boxes for each predictor that approach the appropriate position,
size, and shape of the object. Next for each anchor box, find which object’s
ep
185 bounding box has the biggest IoU (Intersection Over Union) and for IoU more
than 50%, it instructs the neural network to learn the object with greatest
IoU. Else, it notifies the neural network that the real detection is ambiguous
Pr
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
and that the example should not be learnt. If the maximum IoU is less than
iew
40%, the anchor box should assume that there is no object [25]. In practise,
190 this works well, and the hundreds of predictors do a great job of assessing
whether or not an image contains a particular type of object.If default an-
chor box setup is used, it can result in predictors that may not reach a 50%
IoU with any of the anchor boxes. In this instance, the neural network will
never be aware of the existence of these entities and so will never be able to
ev
195 detect them.
r
proach) is used. It is computationally faster and more efficient, to find a
collection of aspect ratios that cover the majority of shapes in the dataset
[26]. This has been accomplished by identifying common clusters of the
er
pe
ot
tn
Figure 5: Height Width Scatter plot and Anchor Box Visualisation from the Training
rin
Dataset
200
dataset’s bounding boxes and then locating the centroids of these clusters
using the k-means clustering. It created thousands of anchor boxes for our
dataset as shown in Figure 5. The visualization of height and width scatter
ep
plot shows that most concentrated values in the lower range. Therefore, for
205 training the model with best suitable anchors and aspect ratio (height/width
ratio for the anchor boxes) values are chosen in order to match one of multiple
Pr
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
pre-defined anchor boxes to the ground truth bounding boxes while training
iew
[27]. In this work, 5, 10, and 15 aspect ratios of anchors for training the
model have been selected and results have been compared with the default
210 anchor value of 3 with aspect ratios 0.5, 1.0, 2.0. Table 1 provides the set of
anchor values and aspect ratio used. For visualization and comparison pur-
pose, anchor boxes for default anchors and 15 anchors are shown in Figure
6.
ev
Anchor Values Aspect Ratios
3 0.50, 1.00, 2.00
5 0.15, 0.75, 1.00, 1.34, 6.90
10 0.10, 0.30, 0.40, 0.50, 0.70, 1.80, 2.30, 2.80, 4.00, 6.90
15 0.07, 0.15, 0.24, 0.35, 0.45, 0.65, 0.75, 1.00,
r
1.34, 1.55, 2.25, 2.90, 4.17, 6.90, 15.40
er
pe
ot
tn
and test subsets. During detection, the model predicts numerous bounding
boxes for each item and eliminates unneeded boxes based on their confidence
value. The performance of the model is computed based on the following
metrics:
Pr
10
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
220 • Accuracy: It is one of the important classification metric that is defined
iew
as ”The total number correct predictions divided by the total number
of predictions, into hundred.”
ev
precision averaged across all distinct recall levels. This parameter is
important to compare different detectors in terms of a numerical met-
ric that could be directly used for comparison.
r
230 • Average Recall (AR): Recall is the number of true positives divided by
the number of ground-truths. It is estimate of how much the model
er
has remembered. Average recall is the recall averaged over all IoUs and
can be computed as two times the area under the recall-IoU curve.
• True Positive (TP): IoU is the measure of overlap between the pre-
pe
235 dicted bounding box and a ground-truth box. For IoU greater than the
threshold, the prediction is classified as TP. Here, a threshold of 50%
(0.5) is used to derive the results.
object is over 50%, only then the label will be taken for evaluation.
245 This section presents the results achieved through the proposed method-
ology described in the previous sections. A set of experiments were per-
formed on the Small Nut Surface Defect Dataset and then the NEU-DET
Dataset. The class wise performance for — scratch, dent, crack, and non-
ep
defectives(ND) of the Small Nut Surface Defect Dataset using 3, 5, 10, and
250 15 anchors is as presented in Figure 7. The model with 15 anchors has pro-
vided greater precision in the cases of dent, crack and ND. In case of scratch,
the greatest precision is provided by model with 10 anchors. Considering
Pr
11
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
scratch as a non-critical defect, it can be concluded that 15 anchors are best
iew
suitable for the small nut surface defect dataset. Also the classification loss
and localization loss is minimum for in case of 15 anchors.
r ev
er
pe
255
have been conducted. The first experiment used the default anchor size of 3,
which has only three aspect ratio values. With an anchor value of 15 and a
260 50% IoU, accuracy is 98.04% from a class-wise classification of the fasteners.
This outperforms the default anchor value of 3 which has an accuracy of
rin
all, while 15 anchor values reported best average accuracy, precision, and
recall. An investigation of the relative performance of all the anchor values
shows that 15 anchors performed best as presented in Figure 8. This further
Pr
12
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
0.985
iew
0.98
0.975
0.97
0.965
0.96
ev
0.955
0.95
0.945
0.94
r
0.935
Anchors=3 Anchors=5 Anchors=10 Anchors=15
Accuracy
er
mAP@0.5 Average Recall
since each weight in the training network will be updated frequently. These
weights are updated in accordance with the learning rate assigned to the
model. Figure 9 shows the predicted results for the proposed model in case
280 of small nut dataset. Figure 10 illustrates the difference between the train-
ing of models without anchor tuning and with anchor tuning. While training
rin
without anchor tuning, it can be seen that the learning rate falls drastically
within 5000 training steps. This means that the weights of the model are
updated divergently. On the other hand, with anchor tuning learning rate
285 varies slowly with respect to the training steps since updates to the weights
ep
of the model are also slower and result in lower loss function. This way the
training extended to 35,000 training steps with learning rate of value 0.0179.
Figure 11 compares the total loss value all over the training steps with and
Pr
13
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
Overall Results
Anchors
mAP @0.5 Accuracy Ave. Recall Tot. Pred.
iew
3 0.95 0.95 0.96 431 (42%)
5 0.96 0.97 0.97 540 (53%)
10 0.97 0.97 0.97 616 (60%)
15 0.97 0.98 0.98 703 (68%)
Classwise Accuracy
Anchors
Dent Crack Scratch Non Defectives
3 0.98 1.00 0.96 0.95
5 0.99 0.99 0.98 0.97
10 0.99 0.99 0.98 0.98
ev
15 0.99 0.99 0.99 0.99
Classwise Precision
Anchors
Dent Crack Scratch Non Defectives
3 0.92 1 0.92 0.99
5 0.97 0.98 0.92 1.00
10 0.97 0.96 0.97 0.99
r
15 0.98 0.94 0.99 0.96
Classwise Recall
Anchors
Dent Crack Scratch Non Defectives
3
5
10
15
Anchors
1.00
0.99
0.99
0.98
Dent
1.00
1.00
0.98
1.00
Crack
er 0.96
0.99
0.98
0.99
F1 Score
Scratch
0.89
0.91
0.95
0.96
Non Defectives
pe
3 0.96 1.00 0.94 0.94
5 0.98 0.99 0.96 0.96
10 0.98 0.97 0.97 0.97
15 0.98 0.97 0.99 0.98
Table 3: Comparison with other models for Small Nut Dataset with Pred. as Prediction
without anchor tuning. It is apparent that using a model without anchor tun-
rin
290 ing causes a larger overall loss. Anchor size has been adjusted based on the
different types of defects identified in the dataset. As a result, the model can
now detect small, large, and irregular faults, which reduces the overall loss
value. Classification loss occurs when the bounding box does not match the
ep
14
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
iew
r ev
(a) Labeled (b) Predicted
er
pe
ot
the predicted bounding box and labeled bounding box. Figure 12 illustrates
the classification loss, regularization loss, and localization loss for various
300 anchor sizes. The minimum value of classification loss of 0.42 is achieved
in case of 3 anchors. The regularization loss remains unaffected by anchor
rin
predictions made by model are then compared with the test set by matching
the class and location of defect present in both the predicted image and test
image. The four entities (true positive, false negative, false positive, true
negative) of confusion matrix have been calculated based on the comparison
Pr
15
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
iew
r ev
er
Figure 10: Comparison of Learning Rate
pe
ot
tn
rin
made.
ep
310
Table 4 presents the comparison of the proposed model with anchor tuning
and without anchor tuning in the case of NEU-DET dataset. With default
anchors having 3 values the precision of 69%, accuracy of 96.24%, 80% recall
Pr
16
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
iew
r ev
er
Figure 12: Loss Values at different Anchor Sizes
pe
ot
tn
is achieved and the prediction rate is only around 25%. On the other hand
315 in the case of model with anchor tuning to train the model. It achieved
precision of 99%, accuracy of 99.82% with the 99% of recall rate, and 67%
Pr
17
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
Model mAP@0.5 Recall F1-Score
FRCN 0.70 0.93 0.80
iew
SSD 0.72 0.96 0.83
YOLOv2 0.50 0.74 0.60
YOLOv3 0.40 0.73 0.52
EEDN 0.72 0.99 0.84
IMN-YOLOv3 0.98 0.96 0.97
Proposed SDDNet 0.99 0.99 0.99
ev
of prediction rate is achieved.
Table 5 shows the detailed comparison results of proposed model with
other models in terms of mAP@0.5, recall and F1 score in the case of NEU-
r
320 DET dataset . In this paper, the test set of the dataset is used to evaluate
proposed model and competitive results with 99% mAP@0.5, 99.8% of recall
er
rate, and achieved 99.3% F1 score in the real-time scenario at 25fps.
5. Conclusion
pe
In this article, single shot detectors (SSD) approach based on MobilenetV2
325 also known as Surface defect detection network (SDDNet) was proposed for
object detection classification and localization specifically fasteners. The
study was also based on the Small Nut Surface Defect Dataset of M6 and
M8 nut sizes and a publicly open dataset. The experimental result shows that
ot
and aspect ratios centered on each pixel. The best result achieved was with
anchor size 15, thus leading to a conclusion as the optimal anchor size for
the small nut surface defect dataset when scratch is taken into consideration.
340 Using SSDNet with anchor tuning provides the advantage of evaluating the
ep
18
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
345 Acknowledgement
iew
This research was supported under the NTU-PU (Nottingham Trent Uni-
versity, United Kingdom and Panjab University, India) Science and Technol-
ogy Partnership Collaborative (STPC) research grant.
References
ev
350 [1] I. Pastor-López, I. Santos, A. Santamarı́a-Ibirika, M. Salazar, J. De-
la Peña-Sordo, P. G. Bringas, Machine-learning-based surface defect
detection and categorisation in high-precision foundry, in: Industrial
Electronics and Applications, 2012, pp. 1359–1364.
r
[2] S. Kim, W. Kim, Y.-K. Noh, F. C. Park, Transfer learning for auto-
355
[3] C. Gudas, et al., The effects of fatigue cracks on fastener loads during
pe
cyclic loading and on the stresses used for crack growth analysis in clas-
sical linear elastic fracture mechanics approaches, Materials Sciences
360 and Applications 11 (2020) 505.
365 [5] E. Westphal, H. Seitz, A machine learning method for defect detec-
tion and visualization in selective laser sintering based on convolutional
neural networks, Additive Manufacturing 41 (2021) 1–13.
19
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
[8] R. Manish, A. Venkatesh, S. D. Ashok, Machine vision based image pro-
iew
375 cessing techniques for surface finish and defect inspection in a grinding
process, Materials Today: Proceedings 5 (2018) 12792–12802.
ev
380 [10] Y. Xu, K. Zhang, L. Wang, Metal surface defect detection using modified
yolo, Algorithms 14 (2021) 1–14.
r
for steel metal surface based on deep learning, in: Information Technol-
ogy and Mechatronics Engineering Conference, 2020, pp. 637–641.
385
er
[12] L. Song, X. Li, Y. Yang, X. Zhu, Q. Guo, H. Yang, Detection of micro-
defects on metal screw surfaces based on deep convolutional neural net-
works, Sensors 18 (2018) 1–14.
pe
[13] M. N. A. A. Patar, M. A. Ayub, N. A. Zainal, M. A. Rosly, H. Lee,
A. Hanafusa, Detection of micro-defects on metal screw surfaces based
390 on faster region-based convolutional neural network, in: Intelligent Man-
ufacturing and Energy Sustainability, Springer, 2022, pp. 587–597.
ot
395
[15] X. Lv, F. Duan, J.-j. Jiang, X. Fu, L. Gan, Deep metallic surface defect
detection: The new benchmark and detection network, Sensors 20 (2020)
1–15.
rin
20
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498
ed
[18] Y. He, K. Song, Q. Meng, Y. Yan, An end-to-end steel surface defect de-
iew
tection approach via fusing multiple hierarchical features, IEEE Trans-
actions on Instrumentation and Measurement 69 (2019) 1493–1504.
[20] Y.-C. Chiu, C.-Y. Tsai, M.-D. Ruan, G.-Y. Shen, T.-T. Lee, Mobilenet-
ev
ssdv2: An improved object detection model for embedded systems, in:
International conference on system science and engineering, IEEE, 2020,
pp. 1–5.
r
415 [21] K. Simonyan, A. Zisserman, Very deep convolutional networks for large-
scale image recognition, arXiv preprint arXiv:1409.1556 (2014).
er
[22] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, A. C.
Berg, Ssd: Single shot multibox detector, in: European conference on
computer vision, Springer, 2016, pp. 21–37.
pe
420 [23] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang,
T. Weyand, M. Andreetto, H. Adam, Mobilenets: Efficient convolu-
tional neural networks for mobile vision applications, arXiv preprint
arXiv:1704.04861 (2017).
ot
430 [26] H. Qiu, H. Li, Q. Wu, F. Meng, K. N. Ngan, H. Shi, A2rmnet: Adap-
tively aspect ratio multi-scale network for object detection in remote
sensing images, Remote Sensing 11 (2019) 1–23.
21
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4254498