Received: 28 February 2023 Revised: 18 May 2023 Accepted: 21 May 2023 IET Generation, Transmission & Distribution

DOI: 10.1049/gtd2.12886

ORIGINAL RESEARCH

Optimized hybrid YOLOu-Quasi-ProtoPNet for insulators

classification

Stefano Frizzo Stefenon1,2 Gurmail Singh3 Bruno José Souza4

Roberto Zanetti Freire5 Kin-Choong Yow6

1
Digital Industry Center, Fondazione Bruno Kessler, Abstract
Trento, Italy
To ensure the electrical power supply, inspections are frequently performed in the power
2
Department of Mathematics, Computer Science grid. Nowadays, several inspections are conducted considering the use of aerial images
and Physics, University of Udine, Udine, Italy
since the grids might be in places that are difficult to access. The classification of the
3
Department of Computer Sciences, University of insulators’ conditions recorded in inspections through computer vision is challenging, as
Wisconsin-Madison, Madison, Wisconsin, USA
object identification methods can have low performance because they are typically pre-
4
Industrial and Systems Engineering Graduate
trained for a generalized task. Here, a hybrid method called YOLOu-Quasi-ProtoPNet is
Program (PPGEPS), Pontifical Catholic University
of Parana (PUCPR), Curitiba, Brazil proposed for the detection and classification of failed insulators. This model is trained
5
Universidade Tecnológica Federal do Paraná
from scratch, using a personalized ultra-large version of YOLOv5 for insulator detection
(UTFPR), Curitiba, Brazil and the optimized Quasi-ProtoPNet model for classification. For the optimization of the
6
Faculty of Engineering and Applied Science, Quasi-ProtoPNet structure, the backbones VGG-16, VGG-19, ResNet-34, ResNet-152,
University of Regina, Regina, Saskatchewan, Canada DenseNet-121, and DenseNet-161 are evaluated. The F1-score of 0.95165 was achieved
using the proposed approach (based on DenseNet-161) which outperforms models of the
Correspondence same class such as the Semi-ProtoPNet, Ps-ProtoPNet, Gen-ProtoPNet, NP-ProtoPNet,
Stefano Frizzo Stefenon, Digital Industry Center,
Fondazione Bruno Kessler, Via Sommarive 18,
and the standard ProtoPNet for the classification task.
Trento, TN 38123, Italy.
Email: sfrizzostefenon@fbk.eu

Funding information
Natural Sciences and Engineering Research Council
of Canada (NSERC), Grant/Award Number:
DDG-2020-00034; Conseil de recherches en
sciences naturelles et en génie du Canada (CRSNG),
Grant/Award Number: DDG-2020-00034

1 INTRODUCTION tion, insulators are vulnerable to vandalism and other issues

because they are mostly installed outdoors [7–9].
Insulators in the transmission lines are components respon- When insulators lose their insulating properties intermit-
sible for supporting the electrical power grid and isolating tent discharges can happen, affecting power quality [10]. With
the electrical potential [1]. When there is an accumulation of increasing partial discharges and leakage current [11], the con-
contamination on the surface of these components, electrical dition can worsen until disruptive discharges occur, resulting in
discharges may occur [2–4]. Contamination is difficult to the shutdown of the power grid [12]. In many cases after a shut-
measure because its presence does not represent an immi- down, the system is re-established and the fault does not have
nent failure, besides the rain can help clean the insulators [5]. the same features (distribution of the contamination over the
Over time the contamination becomes encrusted and strongly insulation surface), which makes it difficult to locate. The elec-
attached to the surface of insulators, causing these components trical outage represents a serious problem for the reliability of
to lose their insulating properties [6]. Apart from contamina- the electricity supply [13], reducing the power quality indices

This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided
the original work is properly cited and is not used for commercial purposes.
© 2023 The Authors. IET Generation, Transmission & Distribution published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology.

IET Gener. Transm. Distrib. 2023;1–11. wileyonlinelibrary.com/iet-gtd 1

2 STEFENON ET AL.

that are used to measure whether the electric utility is adequately computer vision are presented. Section 3 presents the proposed
serving consumers [14]. YOLOu-Quasi-ProtoPNet model. In Section 4 the structure
To mitigate electrical system shutdowns, inspections are car- optimization of the model and fine-tuning are presented and
ried out by specialized teams to identify signs of adverse discussed. In Section 5 the final considerations are presented
conditions [15]. Power system examinations can be performed and further works are suggested.
using specific equipment or visual inspections [16]. The equip-
ment that is generally used in inspections are acoustic detectors
[17], ultraviolet sensors [18], infrared cameras [19], and others 2 RELATED WORKS
[20]. One of the disadvantages of using this specific type of
equipment is that the operator needs to be specialized in its The evaluation of transmission power system insulators through
operation having the ability to interpret a possible indication inspections is important to keep the power grid in good working
of a failure, which is difficult due to the need for multitasking condition [26]. Several authors have been researching state-
operators [21]. of-the-art models to improve network inspections [27–29].
Among the image processing techniques for classification, Although network monitoring is efficient using specific equip-
convolutional neural networks (CNNs) have been widely used ment, in some situations it is unfeasible to take measurements in
for pattern recognition [22], highlighting specific applications hard-to-access places, which can be solved by visual inspections
for electrical power system fault identification [23]. There are [30]. The use of computer vision becomes a promising alterna-
several variations of these models and their applications are tive to obtain a model that is effective in identifying adventitious
promising for the identification of adverse conditions in the conditions in the electrical power grid [31].
power grid. Currently available, there is the prototypical part Besides porcelain insulators, glass insulators, and polymeric
network (ProtoPNet), which as a major advantage has inter- insulators can be found in electrical power transmission lines.
pretability in some cases, beyond improving the classification Glass insulators have similar characteristics to porcelain insu-
can assist in the interpretation of the result. Based on this char- lators, considering that these materials have a high fusion
acteristic, the ProtoPNet is promising to solve the problem of temperature, which results in greater robustness to electric dis-
classification presented here. charges [32], and therefore, they are the most common in
Generally, CNNs have difficulties identifying small objects, electrical networks that are in operation for a long time (over
you only look once (YOLO) stands out to solve this task [24], 30 years). Polymeric insulators have been recently used because
having a better performance than sliding windows methods they are lighter, making them easier to install and perform main-
based on standard CNNs [25]. Considering that the YOLO tenance [33]. The model proposed here could be applied to
model has a high capacity to detect objects and the ProtoP- other types of profiles and materials, beyond those evaluated in
Net approach has a higher classification capacity, here, we this work, being necessary to train the model with the insulators
propose the hybrid optimized YOLOu-Quasi-ProtoPNet that in question.
combines the best advantages of each class of these models. The According to Salem et al. [34], the insulator profile has
contributions of this paper are: an influence on its performance. They have evaluated insula-
tors coated with room-temperature vulcanizing considering the
∙ The first contribution is the object detection performance influence of humidity on flashover based on the evaluation of
improvement of the YOLOv5 model proposing a person- different profiles. Salem et al. [35], using finite element methods
alized ultra-large version (YOLOv5u), which has proven (FEM), presented interesting results regarding the difference of
superior to the variations of the standard YOLOv5 mod- profiles using glass insulators. They have highlighted that the
els (YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l, and issue depends on the location and dimension of the pollution
YOLOv5x). region. FEMs have been applied to evaluate the design of power
∙ The second contribution is related to the use of a hybrid system components by simulating dynamic variations and opti-
approach, which uses the best capabilities of the YOLOv5u mizing the component’s structure [36], making it promising for
for object detection and a Quasi-ProtoPNet model for clas- the definition of the insulator design [37, 38] and its ability to
sification. Considering the evaluation of VGG-16, VGG-19, support stress [39].
ResNet-34, ResNet-152, DenseNet-121, and DenseNet-161 Due to the popularization of unmanned aerial vehicles
as backbone. (UAVs), it is becoming increasingly common to use these types
∙ The third advantage of the proposed model is its explainabil- of aircraft to perform image-based monitoring of the power grid
ity. The explainability helps the users to understand why the [40]. The main advantage of this approach is that UAVs can per-
model performs the classification, helping the maintenance form grid inspection in hard-to-reach places. Field teams that
teams work on the need. The Quasi-ProtoPNet uses proto- travel along the branch line taking photographs have a hard time
types of large spatial dimensions that help the model classify doing this work when there is a large variation in relief, mak-
images based on objects rather than the backgrounds of the ing the inspection unfeasible in some situations. For this reason,
objects in the images. the use of UAVs is a promising alternative for power system
inspection [41].
The rest of this paper is organized as follows: In Sec- According to Foudeh et al. [42], one of the major difficul-
tion 2 related works regarding insulator fault detection using ties in performing inspections in electrical power systems is
STEFENON ET AL. 3

the lack of adaptability to adverse conditions, thus the use of presented in Figure 1, which is divided into steps from A to F
UAVs has received much attention for overhead electric power and is explained in detail in this section.
lines patrol process. Due to the aging of composite insula- To improve the proposed method, an optimization of the
tors, the phenomenon of micro-cracks occurs, which can be network structure is performed by changing its classifier, in
identified with UAVs through image analysis. Jin et al. [43] which VGG-16, VGG-19, ResNet-34, ResNet-152, DenseNet-
show that through image pre-processing techniques it is pos- 121, and DenseNet-161 are evaluated. In Figure 1a, the images
sible to extract the texture of micro-cracks and identify them of the dataset are loaded, and the detection of insulators is per-
easily. formed using YOLOv5, wherein the structure size variations
Another strategy that can be applied to improve the are evaluated (YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l,
classification of adherent conditions in insulators is image pre- YOLOv5x, and YOLOv5u).
processing. In many situations, filters can be used to take the
focus of the analysis of the background of the image and con-
centrate the evaluation on the fault locations. Edge detection 3.1 You only look once
techniques such as Sobel and Canny edge detection can be used
to highlight the presence of faults, thus improving classification YOLO is a CNN that splits images into grids, having each grid
without representing a significant computational effort increase cell detect objects within itself [51]. This approach is a single-
[44]. shot algorithm, which means that it only requires processing
The fusion convolutional network (FCN) is an outstanding the image once to detect and classify the object under consid-
model for real-time monitoring of the power grid via UAVs. eration. Recently the YOLO has shown satisfactory results, and
The FCN model proposed by Mussina et al. [45] consists of the framework has been improved several times since its initial
a CNN combined with a binary classifier multilayer neural net- release [52].
work. By using the FCN model for a multi-modal information The input images are divided into an S × S grid, where
fusion system, the image classification output of the CNN can each square of the grid (i) predicts the target bounding box,
be combined with the leakage current values, thus obtaining which corresponds to its degree of confidentiality. Therefore,
a model with high classification capability. Among the CNNs, the confidence of the classes (cl ) of objects (ob j ), is given by:
models based on VGG-16 [46], ResNet-101 [47], and AlexNet
[48] have been used for this task. pr (cli |ob j ) ⋅ pr (ob j ) ⋅ IoU pred
truth
= pr (cli ) ⋅ IoU pred
truth
. (1)
The models that have stood out for object detection are the
YOLOs because their architecture is based on a single shot After splitting the image into grids a class probability mean
and generally has better performance results than other models average precision (mAP) is created to identify the target objects
for the same application [49]. With this advantage, this model and bounding boxes to determine if the desired objects are
proves to be promising for the task being presented here. How- located in this confidence region. The YOLOv5 backbone has
ever, since the YOLO models are based on a standard backbone its structure (presented in Figure 1b), different from previous
classifier that in versions YOLOv3 and YOLOv4 are based on versions of the model which use the Darknet-53, this makes this
Darknet53, it may have low classification performance when version more effective and flexible to be modified according to
few images are used for training. To improve this problem, a the needs of the project.
hybrid method that uses another classifier may be an alterna- The YOLOv5 framework has features that makes it promis-
tive to improve the fault identification capability of this class ing for fault classification in insulators. The YOLOv5 applies
of models. Mosaic to improve the detection performance of small objects,
For image classification tasks, often the reasoning can be which is a difficulty when inspections of insulators are carried
based on prototypical aspects of a class. Evidence of this dif- out, considering that the failure may be small in relation to the
ference aids the final decision [50]. The ProtoPNet dissects insulator chain.
the image by finding prototypical parts and combines evidence Using Mosaic, an increase in the data amount is intro-
from the prototypes to make a classification. Based on this duced into the network training process, which consequently
advantage, the combination of the YOLO model for object allows the increase of the batch size, causing each iteration
detection with a model based on ProtoPNet becomes a promis- to have more data. This feature might be an advantage in
ing proposal for the challenge presented here, and this model terms of having additional information to the model, how-
will be presented in more detail in the next section. ever, it results in an increase in the need for data processing
power. Since the training is performed offline, the higher
computation effort in network training does not represent a
3 YOLOu-QUASI-ProtoPNet disadvantage. The application of Mosaic reduces the size of
the target and therefore increases the detection efficiency of
Considering the high performance of the YOLOv5 for object smaller objects, which is a requirement to obtain applicable
detection and the high classification capability of the Quasi- results [53].
ProtoPNet model, here a hybrid model called YOLOu-Quasi- YOLOv5 employs the Focus, BottleneckCSP, SPP, and path
ProtoPNet is proposed. The structure of the proposed model is aggregation network (PANet) techniques to improve object
4 STEFENON ET AL.

(c)
(f)
(b)
(e)

(a) (d)

FIGURE 1 Structure of the proposed YOLOu-Quasi-ProtoPNet.

detection. Focus is applied to improve the receptive field, Bot- since the training is performed offline, and the model has high
tleneckCSP extracts the information from features, and SPP speed in the testing phase. To create the YOLOv5u, the size
separates the features that are relevant to improve the non- of the net was changed and evaluated based on the variation
linear representation of the network. PANet combines the high of depth and width multiple, which is the variation performed
and low-level features to enhance the accuracy detection, PANet in other YOLOv5 versions. The cross-entropy loss function
improvement is highlighted in red in Figure 1c. is defined to calculate the score loss in object detection [56].
For better accuracy, it is common to scale a baseline detec- YOLOv5u creates a bounding box of the detected insulator
tor using a bigger backbone network. The EfficientDet method (Figure 1d) that is validated by its confidence, and the crop-
[54], applied in YOLOv5, uses a bi-directional weighted fea- outs from it are made, these are the outputs of the YOLOv5u.
ture pyramid network (BiFPN) for multiscale feature fusion and After object detection, the outputs are classified by the Quasi-
a composite scaling strategy that uniformly scales the resolu- ProtoPNet model. For CNN models, a batch size multiple of 8
tion of the network, hence it jointly scales all dimensions of was used. The batch sizes were dependent on the base model,
the backbone. The multiple depth is responsible for the depth the heavier the base model the smaller the batch size. VGG-
of the model, meaning that it ends up adding more layers to 16, VGG-19, ResNet-34: batch size = 48, DenseNet-121: batch
the net, whereas the multiple width adds more filters to the size = 32, DenseNet-161: batch size = 24, ResNet-152: batch
layers, thus it adds more channels to the outputs of the lay- size = 16.
ers. The depth_width parameters are responsible for defining the
size of the network and creating the variations of the YOLOv5
models, these parameters will be evaluated here for proposing a 3.2 Quasi-ProtoPNet
custom model.
The proposed personalized ultra-large version of YOLOv5, Quasi-ProtoPNet is a model based on the ProtoPNet approach
named YOLOv5u emerged from the consideration that larger that uses prototypes to simulate human reasoning. Specifically,
models have better mAP results in preliminary experiments pre- Quasi-ProtoPNet uses only a positive reasoning process, plac-
sented by Ultralytics.1 A progression in classification capability ing a zero binding across similarity scores and misclassifications.
is observed when increasing network size, for this reason, a Quasi-ProtoPNet does not perform convex optimization of the
model larger than YOLOv5x seems promising, with YOLOv5x last layer to maintain constant connections, in other words, the
having the best results compared to other models with fewer model does not freeze all other layers to optimize the last dense
parameters such as YOLOv5n, YOLOv5s, YOLOv5m, and layer [57].
YOLOv5l. The results of these models are based on a pre- Besides the positive reasoning process, Quasi-ProtoPNet
trained dataset using the common objects in context (COCO) employs prototypes of all spatial dimensions, meaning rect-
[55]. Here, all these models were trained from scratch to have a angular and square spatial dimensions, while the ProtoPNet
fair comparison. models generally use prototypes having only square spatial
The disadvantage of using an ultra-large model is that it dimensions. Other models based on ProtoPNet are increas-
requires more powerful hardware, and has higher FLOPs, an ingly being used, most notable are Semi-ProtoPNet [58],
approach that results in higher computational effort and more Ps-ProtoPNet [59], Gen-ProtoPNet [60], and NP-ProtoPNet
time needed for training. However, this is not a problem here, [61]. The Gen-ProtoPNet uses a generalized version of the
Euclidean distance function, the NP-ProtoPNet considers
1
https://github.com/ultralytics/yolov5 the negative reasoning process and the positive reasoning
STEFENON ET AL. 5

process, but it emphasizes the negative reasoning process, and

the Ps-ProtoPNet equally considers both types of reasoning
processes and uses fixed connections between similarity scores
and logits.
Quasi-ProtoPNet has convolution layers of a base model
that are followed by two additional convolution layers, here
the VGG-16, VGG-19, ResNet-34, ResNet-152, DenseNet-
121, and DenseNet-16 are used as baselines. The convolutional
layers are denoted by L, and are followed by a generalized con-
volutional layer pt of prototypical parts. The pt layer is followed
by a dense w layer, without bias. The weight matrix of the dense
layer is, respectively, denoted by wm . In this structure, the L
convolutional layers are the non-interpretable part of the Quasi-
ProtoPNet (Figure 1e), while the pt forms the interpretable part
of the model (Figure 1f).
Quasi-ProtoPNet employs the generalized version of the
Euclidean distance function (d ). Since p is any prototype with
the shape 512 × h × w, wherein 1 ≤ h, w ≤ 6, and h and w
together are not equal to 1 nor 6. The model output (= L(x))
of L has (7 − h)(7 − w) patches of dimensions h × w. There-
fore, the square of the distance is d (i j , p) between p and (i, j )
patch i j of  is given by: FIGURE 2 Annotated images from the used dataset [63].

∑
h ∑
w ∑
512
where the cluster cost (ClstCost) is given by:
d 2 (i j , p) = ||(i+l −1)( j +m−1)k − plmk ||22 . (2)
l =1 m=1 k=1
1∑
n

If p has prototypes of spatial dimensions 1 × 1 (h = w = 1), ClstCost = min min d 2 (, p j ). (8)
n i=1 j ∶p j ∈Pyi  ∈ patches(L(xi ) )

∑
512
Considering x is an input image, the model projects proto-
d 2 (i j , p) = ||i jk − p11k ||22 , (3) types over patches of x which are more similar to the prototypes
k=1
[62]. Therefore, a patch of x is projected that is at a smaller
which is the Euclidean distance square between p and a patch of distance from a prototype, given the following update:
, in which p11k ≃ pk . The prototypical unit pt computes
pcj ⟵ arg min d (, pcj ).
( ) {∶ ∈ patches(L(xi ) ) ∀i such that yi =c}
d 2 (i j , p) + 1
pt () = max log . (4)
1≤i≤7−h, 1≤ j ≤7−w d 2 (i j , p) + 𝜖
3.3 Dataset
Thus,
The used dataset was created by Lewis and Kulkarni [63], for
( ) a competition with the goal of insulator defect detection. The
d 2 (, p) + 1
pt (L(x)) = max log . (5) dataset has high-quality labeled images of transmission line insu-
 ∈ patches(L(x) ) d 2 (, p) + 𝜖 lators, which contain four classes: insulators with flashover,
broken insulators, good insulators, and insulator chains. Since
The Quasi-ProtoPNet is trained using two steps, the opti- the purpose of this paper is to classify the condition of insula-
mization of all layers before the dense layers, and the projection tors (faulty and good), the insulator chain class was disregarded.
of prototypes [57]. Given X = {x1 … xn } and Y = {y1 … yn } are, Some of the insulators from this database are shown in Figure 2.
respectively, sets of images and labels, and The database contains high-resolution pictures of porcelain
insulator chains from power lines, on which the positions and
D = {(xi , yi ) ∶ xi ∈ X , yi ∈ Y }, (6) classes of insulators are noted according to bounding boxes
regarding the vertical and horizontal position and their condi-
the objective function of the Quasi-ProtoPNet is: tion. The images were recorded using digital single-lens reflex
cameras (DSLRs) during aerial inspections of the electrical
1∑
n power grid. The DSLRs settings were adjusted according to
min CrosEnt(h◦pt ◦L(xi ), yi ) + 𝜆ClstCost, (7) the need for brightness compensation, considering the use of
P, Lconv n i=1
the Canon PowerShot G10, Nikon D810, and Nikon D90.
6 STEFENON ET AL.

The dataset has 1596 images, wherein 1195 were used dur- ate the computational effort, floating-point operations (FLOPs)
ing the training, and 401 for testing. The high-resolution are used.
images are rescaled to 640 pixels on input to the model to For the object detection task, the nano, small, medium, large,
have standard images according to the training needs of the extra-large, and ultra-large versions of the YOLOv5, of which
presented architecture. are YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l, YOLOv5x,
The cutouts considered for the classification task were 10,886 and YOLOv5u are evaluated. These versions were based on
for training and 1941 for testing, divided as follows: Broken 640-pixel images and the difference between them is the num-
class (train: 877, test: 191), flashover class (train: 1651, test: 288), ber of layers and parameters. These variations in size are given
and good class (train: 8358, test: 1462). The cutouts have varying by the depth and width multiple of the network. Especially
resolutions according to the bounding boxes of the insulator. the ultra-large version is not available in Ultralytics models, the
After augmented data, the dataset has thirty times the origi- YOLOv5u was created based on increasing both the layers and
nal number of images (broken: 26,310, flashover: 49,530, good: the parameters of the network, this is a customized version
250,740). created here for comparison to the standard models.
All compared models presented here were trained from
scratch to have an equivalent condition for comparison. Since
3.4 Experiment setup the models were trained from scratch a maximum value of
10,000 epochs for training was defined, considering that in the
The purpose of the present experiment is to compare object initial experiments, the models converged before 1000 epochs,
detection and classification models to determine the best frame- to improve the development of the analysis the early stopping
work for identifying insulator failures in transmission grids. patience was used, in which the model ends the training if there
The measures of object detection performance here were pre- are 50 epochs without improvement.
cision, recall, F1-score, and mAP. The positive predictive value
(precision) and the sensitivity (recall) are given by:
4 RESULTS AND DISCUSSION
tp
Precision = , (9)
t p + fp Whereas the proposed model uses a personalized ultra-large
tp version for object detection (YOLOv5u), the first analysis step
Recall = . (10) is to define the appropriate structure size of the network. At this
t p + fn
stage of the analysis, all annotated insulators are considered to
These measures are obtained from the confusion matrix be of the same class, in view that YOLOv5u is used for object
considering the true positive (t p), false positive ( fp), and false detection, and Quasi-ProtoPNet is used specifically for classifi-
negative ( fn). From the precision and recall, the F1-score is cation. The results of object detection concerning the variation
obtained, according to: in the size of the network are presented in Table 1.
Since YOLOv5u is trained from scratch, it is needed to set
2 × Precision × Recall a higher maximum value of epochs, 10,000 here, which is con-
F1−score = . (11)
Precision + Recall siderably more than the standard 300 epochs suggested in the
model’s original repository. Considering the used early stop cri-
The intersection over union (IoU) determines when detec- teria, the compared models converged at close to 1000 epochs.
tion is considered a true positive. A detection of t p is defined by Here, the best results of comparative analyzes between models
IoU > T , in which T is a predefined threshold. Here, the valua- of the same class are underlined and the best overall results are
tion is made by T equal to 0.5. To evaluate the object detection, highlighted in bold.
the mAP was used, given by: In this initial comparison, all the models were able to achieve
F1-score results above 0.97, which shows that the approach is
1∑
n
mAP = AP k , (12) feasible for the application in question. As expected, the models
n k=1 that have more parameters need more time to be computed due
to the higher computational effort required, which can also be
where n is the number of classes and k is the corresponding observed by the higher value of FLOPs.
class. The average precision (AP) calculates a precision–recall The mAP@[0.5] was greater than 0.98 in all compared mod-
curve as the weighted average of the precision achieved for els, showing that object detection is successfully performed
each threshold. using YOLOv5 in all its versions. Considering the F1-score
The method proposed here was implemented using Python result the model that presented the best result was YOLOv5u
language. To perform the training and comparative analysis, a (1.50_1.67) using 465 layers and 158.5 million parameters. For
cluster was used, in which the following setup requirements this reason, this structure has been defined as the standard
were allocated: A graphic processing unit (GPU) GeForce GTX model for object detection, since at this stage of the analysis
1080 Ti, and 64 GB of random access memory (RAM). This all the insulators are considered equal because the objective of
hardware setup was used since it was sufficient to compute the YOLOv5 model is to locate the components for further
all the experiments and thus facilitate reproducibility. To evalu- classification by the Quasi-ProtoPNet model.
STEFENON ET AL. 7

TABLE 1 Definition of structure for object detection.

mAP
Parameters FLOPs
Model depth_width Layers (M) (G) Time (s) Precision Recall F1-score [0.5] [0.5:0.95]

YOLOv5n 0.33_0.25 213 1.9 4.5 27,884.0 0.98092 0.96806 0.97445 0.98757 0.84610
YOLOv5s 0.33_0.50 213 7.2 16.5 28,477.8 0.97797 0.97424 0.97610 0.98720 0.88088
YOLOv5m 0.67_0.75 290 21.2 49.0 38,551.6 0.97949 0.97372 0.97659 0.98492 0.89924
YOLOv5l 1.00_1.00 367 46.5 109.1 60,093.9 0.97526 0.97527 0.97526 0.98483 0.91291
YOLOv5x 1.33_1.25 444 86.7 205.7 74,945.3 0.98120 0.96815 0.97463 0.98271 0.91187
YOLOv5u 1.33_1.50 444 124.1 293.0 133,823.4 0.98110 0.97269 0.97688 0.98353 0.91191
1.50_1.50 465 127.4 307.2 139,318.1 0.97774 0.97311 0.97542 0.98333 0.91234
1.50_1.67 465 158.5 384.6 182,122.8 0.98057 0.97476 0.97765 0.98314 0.90545
1.67_1.67 521 179.7 431.9 190,239.1 0.97411 0.97476 0.97443 0.98260 0.91248
1.67_1.75 521 196.6 469.0 272,763.3 0.97375 0.97372 0.97373 0.98420 0.91679

The difference between the YOLOv5u (1.50_1.67) which

had the best F1-score result to the YOLOv5n, which
uses fewer parameters, was 0.0032. YOLOv5n had the best
mAP@[0.5] results, lower time to be computed, and higher
efficiency considering the FLOPs, however, it had the low-
est mAP@[0.5:0.95] value. This shows that depending on the
research objective, smaller models can have an acceptable F1-
score and even be better in some circumstances relative to
mAP@[0.5].
The time required for training increases when the model uses
more parameters, the difference between the model that used
a depth_width equal to 1.67_1.75 to the YOLOv5n was approx- FIGURE 3 Testing images results: (a) Broken insulator; (b) Flashover
imately 10 times more time to complete the training. Although insulator.
the training time was higher, all models needed less than 3 ms to
process each image during the testing phase. This makes their
application in embedded systems promising since in the test- ProtoPNet is a good strategy for the evaluation in question.
ing phase the computational effort may be limited and a fast Figure 3 shows examples of results with inference images (not
response is required. used in training) using the proposed method.
The output of the YOLOv5u model provides the posi- In Figure 3a, a broken insulator is presented, and in Figure 3b
tion of the insulator and based on an image cutout of where an insulator with a flashover over its surface is presented. These
the insulator is, the classification of the component is per- failures are the most common to be identified in a power
formed. To perform an optimization on the structure of the grid inspection and are the focus of the application of the
Quasi-ProtoPNet, the VGG-16, VGG-19, ResNet-34, ResNet- proposed method.
152, DenseNet-121, and DenseNet-161 baselines are evaluated.
Table 2 shows the classification results using the Quasi-
ProtoPNet model considering this variation, and compares with 4.1 Comparison to related studies
well-established models. In this evaluation, mAP values are not
presented as this metric is based on the IoU of object detec- Serikbay et al. [64] using a starting CNN had an accuracy of
tion that was previously presented in Table 1 for the YOLOv5u 0.8907 in the testing phase. In this application only 1.38 MB of
model. Therefore, Quasi-ProtoPNet is focused specifically on memory was required, making this a promising solution to be
the classification task. applied in an embedded system. Alahyari et al. [65] used a two-
The Quasi-ProtoPNet is superior to all compared models stage model for both the segmentation and the detection tasks
using all backbone variations for the classification task pre- of faulty insulators. The segmentation model achieved a total
sented here. Using the DenseNet-161 as the backbone, the of 78% accuracy, while the classifier obtained 92%. When the
Quasi-ProtoPNet had an F1-score of 0.95165 being promising data is unbalanced the accuracy may not be enough to determine
for the classification of insulators. These results confirm that whether the model is having an acceptable classification result.
using Quasi-ProtoPNet from the object detection performed Zhang et al. [66] used a Fast R-CNN network to identify insu-
by YOLOv5u via a hybrid method defined as YOLOu-Quasi- lator strings. Regarding the detection of the insulator strings, the
8 STEFENON ET AL.

TABLE 2 Comparison of the Quasi-ProtoPNet with different baselines to well-established models.

Quasi- Semi- Ps- Gen- NP-

ProtoPNet ProtoPNet ProtoPNet ProtoPNet ProtoPNet ProtoPNet
Base Metric [57] [58] [59] [60] [61] [62]

VGG-16 Precision 0.89552 0.67187 0.68749 0.56870 0.712499 0.75590

Recall 0.85714 0.23369 0.34591 0.30483 0.31232 0.28318
F1-score 0.87591 0.34677 0.46025 0.39691 0.43428 0.41201
VGG-19 Precision 0.95628 0.51464 0.71249 0.57419 0.57142 0.57516
Recall 0.86206 0.38317 0.31232 0.26567 0.31901 0.25882
F1-score 0.90673 0.43928 0.43428 0.36326 0.40944 0.35699
ResNet-34 Precision 0.95767 0.89893 0.88324 0.58139 0.85572 0.58778
Recall 0.88292 0.75446 0.77678 0.26041 0.77477 0.18075
F1-score 0.91878 0.82038 0.82660 0.35971 0.81323 0.27648
ResNet-152 Precision 0.94764 0.77725 0.74774 0.744075 0.76146 0.74999
Recall 0.89603 0.72246 0.75113 0.69162 0.73127 0.73008
F1-score 0.92111 0.74885 0.74943 0.71689 0.74606 0.73991
DenseNet-121 Precision 0.95897 0.89839 0.87570 0.77102 0.78095 0.71098
Recall 0.92118 0.74666 0.68281 0.72368 0.73214 0.57209
F1-score 0.93968 0.81553 0.76732 0.74660 0.75576 0.63402
DenseNet-161 Precision 0.96891 0.74404 0.68965 0.65868 0.68452 0.56544
Recall 0.93499 0.43554 0.42105 0.37414 0.40069 0.41221
F1-score 0.95165 0.54945 0.52287 0.47722 0.50549 0.47682

algorithm achieved an AP of 91.75% and a recall of 98%, and

accuracy of 98% for the classification task. Sadykova et al. [49]
used YOLOv2 to identify the string of insulators, and then a
classification model to analyze the insulator surface conditions.
The YOLOv2 model achieved an mAP above 98% for detect-
ing insulator strings and an F1-score of 0.95165. Showing that
even the most previous versions can be successfully used for the
presented task. Comparatively, the proposed method presented
here achieved similar values of F1-score and mAP, however
using high-resolution images.
In reference [53], an mAP of 0.99262 was achieved con-
sidering only the task of insulator identification, based on the
ResNet-18 classifier they have an F1-score result of 0.96216.
Based on ResNet-34 in reference [15] an accuracy of 0.9979 and
an F1-score of 0.9964 was achieved for a similar task. As can
be verified, authors who have used CNN-based models have
had promising results in fault identification, with the choice of
model and its fine-tuning depending on the used dataset.

4.2 Limitations FIGURE 4 False positive classification samples.

Quasi-ProtoPNet gives better performance than the series of

ProtoPNet models when classification is to be made over only images over only a few classes. Therefore, this model can be
a few classes. If the number of classes is large then the perfor- really useful for such situations.
mance of the other ProtoPNet models may be better than the In the interpretable results, there were cases where Quasi-
Quasi-ProtoPNet. However, there are many cases similar to the ProtoPNet highlighted the variation in brightness intensity for
case of insulators discussed here when it is needed to classify good insulators, these false positives are presented in Figure 4.
STEFENON ET AL. 9

Considering that these images were classified as insulators in reference number DDG-2020-00034. Cette recherche a été
good condition the interpretability should not be considered for financée par le Conseil de recherches en sciences naturelles et
non-faulty insulators. en génie du Canada (CRSNG), numéro de référence DDG-
The final result has the accumulated error of the identifica- 2020-00034. The author Roberto Z. Freire would like to thank
tion based on YOLOv5u, therefore for the method to have the National Council for Scientific and Technological Develop-
acceptable results, it needs to have the identification of the com- ment (CNPq) of Brazil (grant: 312688/2021-0) for the financial
ponents properly, since the classifier will not be able to classify support of this research.
failures if the component is not identified.
CONFLICT OF INTEREST STATEMENT
The authors declare no conflict of interest.
5 FINAL CONSIDERATIONS
DATA AVAILABILITY STATEMENT
The proposal of using Quasi-ProtoPNet as a classifier instead The data are available upon request to the authors.
of using only YOLO had rewarding results, comparatively,
Quasi-ProtoPNet was superior in using different baselines to ORCID
all other compared models based on ProtoPNet, additionally, Stefano Frizzo Stefenon https://orcid.org/0000-0002-3723-
YOLOv5u showed promise for object detection, hence the pro- 616X
posed hybrid optimized YOLOu-Quasi-ProtoPNet excelled in
all comparisons for solving the problem presented here. REFERENCES
The use of YOLOv5 specifically for detecting insulators had 1. Dong, B., Hu, Y., Yin, F., Jiang, X.: Simulation calculation of 3D electric
superior results to its use for identifying different conditions, field and natural flashover analysis of ice-covered silicone rubber insulator.
this means that the task of detecting these components may be IET Gener. Transm. Distrib. 17(5), 1166–1178 (2023)
2. Mohammadnabi, S., Rahmani, K.: Influence of humidity and contamina-
simpler than their classification. This result is reflected in electri-
tion on the leakage current of 230-kV composite insulator. Electr. Power
cal grid inspections, where the insulator chain is easily identified, Syst. Res. 194, 107083 (2021)
even by people without specific training, but the identification 3. Salem, A.A., Abd-Rahman, R., Ishak, M.T.B., Lau, K.Y., Abdul-Malek, Z.,
of a failure requires advanced training and knowledge of the grid Al-ameri, S., et al.: Influence of contamination distribution in character-
conditions. Numerically, all variations of the YOLOv5 model izing the flashover phenomenon on outdoor insulator. Ain Shams Eng. J.
102249 (2023)
had an F1-score higher than 0.97 and an mAP@[0.5] higher
4. Salem, A.A., Lau, K.Y., Abdul-Malek, Z., Al-Gailani, S.A., Tan, C.W.:
than 0.98 for insulator detection, while the best results of these Flashover voltage of porcelain insulator under various pollution distri-
models were an F1-score of 0.93986 and an mAP@[0.5:0.95] butions: Experiment and modeling. Electr. Power Syst. Res. 208, 107867
of 0.87312 (depth_width of 1.33_1.50) when it was necessary to (2022)
identify the insulator condition, making it clear that the use of a 5. Stefenon, S.F., Corso, M.P., Nied, A., Perez, F.L., Yow, K.C., Gonzalez,
G.V., et al.: Classification of insulators using neural network based
hybrid method for classification is a promising strategy.
on computer vision. IET Gener. Transm. Distrib. 16(6), 1096–1107
Considering that several state-of-the-art models have been (2022)
evaluated and combined, in future work, it would be promising 6. Stefenon, S.F., Bruns, R., Sartori, A., Meyer, L.H., Ovejero, R.G., Leithardt,
to employ these models in embedded systems, where a spe- V.R.Q.: Analysis of the ultrasonic signal in polymeric contaminated insula-
cific device could indicate fault conditions to a maintenance tors through ensemble learning methods. IEEE Access 10, 33980–33991
(2022)
team in the field. This would support the correction of failures
7. Medeiros, A., Sartori, A., Stefenon, S.F., Meyer, L.H., Nied, A.: Com-
that occur suddenly, thereby indicating adverse conditions for parison of artificial intelligence techniques to failure prediction in
a team to identify the failure and correct it more dynamically contaminated insulators based on leakage current. J. Intell. Fuzzy Syst.
when a shutdown occurs. The YOLO model is being explored 42(4), 3285–3298 (2022)
and improved over time, new versions have been made avail- 8. Salem, A.A., Abd-Rahman, R., Rahiman, W., Al-Gailani, S.A., Al-Ameri,
S.M., Ishak, M.T., et al.: Pollution flashover under different contamination
able such as YOLOv6, YOLOv7, and more recently YOLOv8.
profiles on high voltage insulator: Numerical and experiment investigation.
Based on the promising results of using models larger than the IEEE Access 9, 37800–37812 (2021)
extra-large version of YOLO, it becomes interesting to further 9. Sopelsa Neto, N.F., Stefenon, S.F., Meyer, L.H., Bruns, R., Nied, A., Seman,
increase the number of parameters of the model to obtain an L.O., et al.: A study of multilayer perceptron networks applied to clas-
ultra-large version of the latest YOLO structures. sification of ceramic insulators using ultrasound. Appl. Sci. 11(4), 1592
(2021)
10. Zhang, D., Zhang, Z., Jiang, X., Zhang, W., Zhao, J., Bi, M.: Influence
AUTHOR CONTRIBUTIONS of fan-shaped non-uniform pollution on the electrical property of typi-
Stefano Stefenon: writing - original draft. Gurmail Singh: cal type HVDC insulator and insulation selection. IET Gener. Transm.
formal analysis, methodology, software. Bruno Souza: data cura- Distrib. 10(14), 3555–3562 (2016)
tion, software. Roberto Freire: supervision. Kin-Choong Yow: 11. Salem, A.A., Lau, K.Y., Abdul-Malek, Z., Zhou, W., Al-Ameri, S., Al-
supervision. Gailani, S.A., et al.: Investigation of high voltage polymeric insulators
performance under wet pollution. Polymers 14(6), 1236 (2022)
12. Deb, S., Ray Choudhury, N., Ghosh, R., Chatterjee, B., Dalai, S.: Short
ACKNOWLEDGEMENTS time modified hilbert transform-aided sparse representation for sensing of
We acknowledge the support of the Natural Sciences and overhead line insulator contamination. IEEE Sens. J. 18(19), 8125–8132
Engineering Research Council of Canada (NSERC), funding (2018)
10 STEFENON ET AL.

13. Branco, N.W., Cavalca, M.S.M., Stefenon, S.F., Leithardt, V.R.Q.: Wavelet 33. Salem, A.A., Lau, K.Y., Abdul-Malek, Z., Mohammed, N., Al-Shaalan,
LSTM for fault forecasting in electrical power grids. Sensors 22(21), 8323 A.M., Al-Shamma’a, A.A., et al.: Polymeric insulator conditions estima-
(2022) tion by using leakage current characteristics based on simulation and
14. Wang, Y., Huang, L., Shahidehpour, M., Lai, L.L., Zhou, Y.: Impact of experimental investigation. Polymers 14(4), 737 (2022)
cascading and common-cause outages on resilience-constrained optimal 34. Salem, A.A., Lau, K.Y., Rahiman, W., Al-Gailani, S.A., Abdul-Malek, Z.,
economic operation of power systems. IEEE Trans. Smart Grid 11(1), Abd-Rahman, R., et al.: Pollution flashover characteristics of coated insu-
590–601 (2020) lators under different profiles of coating damage. Coatings 11(10), 1194
15. Singh, G., Stefenon, S.F., Yow, K.C.: Interpretable visual transmission lines (2021)
inspections using pseudo-prototypical part network. Mach. Vis. Appl. 34, 35. Salem, A.A., Abd-Rahman, R., Rahiman, W., Al-Gailani, S.A., Al-Ameri,
41 (2023) S.M., Ishak, M.T., et al.: Pollution flashover under different contamination
16. Corso, M.P., Perez, F.L., Stefenon, S.F., Yow, K.C., García-Ovejero, R., profiles on high voltage insulator: Numerical and experiment investigation.
Leithardt, V.R.Q.: Classification of contaminated insulators using k-nearest IEEE Access 9, 37800–37812 (2021)
neighbors based on computer vision. Computers 10(9), 112 (2021) 36. Stefenon, S.F., Seman, L.O., Pavan, B.A., Ovejero, R.G., Leithardt, V.R.Q.:
17. Park, K.C., Motai, Y., Yoon, J.R.: Acoustic fault detection technique for Optimal design of electrical power distribution grid spacers using finite
high-power insulators. IEEE Trans. Ind. Electron. 64(12), 9699–9708 element method. IET Gener. Transm. Distrib. 16, 1865–1876 (2022)
(2017) 37. Stefenon, S.F., Neto, C.S.F., Coelho, T.S., Nied, A., Yamaguchi, C.K., Yow,
18. Kim, Y., Shong, K.: The characteristics of UV strength according to corona K.C.: Particle swarm optimization for design of insulators of distribution
discharge from polymer insulators using a UV sensor and optic lens. IEEE power system based on finite element method. Electr. Eng. 104, 615–622
Trans. Power Delivery 26(3), 1579–1584 (2011) (2022)
19. Jin, L., Tian, Z., Ai, J., Zhang, Y., Gao, K.: Condition evaluation of 38. Stefenon, S.F., Americo, J.P., Meyer, L.H., Grebogi, R.B., Nied, A.: Analy-
the contaminated insulators by visible light images assisted with infrared sis of the electric field in porcelain pin-type insulators via finite elements
information. IEEE Trans. Instrum. Meas. 67(6), 1349–1358 (2018) software. IEEE Lat. Am. Trans. 16(10), 2505–2512 (2018)
20. Dong, M., Wang, B., Ren, M., Zhang, C., Zhao, W., Albarracín, R.: 39. Dong, G., Li, Q., Liu, T., Gao, H., Zhang, M.: Finite-element analy-
Joint visualization diagnosis of outdoor insulation status with optical sis for surface discharge on polyimide insulation in air at atmospheric
and acoustical detections. IEEE Trans. Power Delivery 34(4), 1221–1229 pressure under pulsed electrical stress. High Voltage 5(2), 166–175
(2019) (2020)
21. Stefenon, S.F., Freire, R.Z., Meyer, L.H., Corso, M.P., Sartori, A., Nied, A., 40. Tang, Z., Jia, C., Wang, H., Rong, S., Zhao, W.: Intelligent height measure-
et al.: Fault detection in insulators based on ultrasonic signal processing ment technology for ground encroachments in large-scale power transmis-
using a hybrid deep learning technique. IET Sci. Meas. Technol. 14(10), sion corridor based on advanced binocular stereovision algorithms. IET
953–961 (2021) Gener. Transm. Distrib. 17(2), 448–460 (2023)
22. Xu, Y., Yan, X., Sun, B., Zhai, J., Liu, Z.: Multireceptive field denois- 41. Yang, L., Fan, J., Liu, Y., Li, E., Peng, J., Liang, Z.: A review on state-of-the-
ing residual convolutional networks for fault diagnosis. IEEE Trans. Ind. art power line inspection techniques. IEEE Trans. Instrum. Meas. 69(12),
Electron. 69(11), 11686–11696 (2022) 9350–9365 (2020)
23. Tao, X., Zhang, D., Wang, Z., Liu, X., Zhang, H., Xu, D.: Detection of 42. Foudeh, H.A., Luk, P.C.K., Whidborne, J.F.: An advanced unmanned aerial
power line insulator defects using aerial images analyzed with convolu- vehicle (UAV) approach via learning-based control for overhead power
tional neural networks. IEEE Trans. Syst. Man Cybernet. 50(4), 1486–1498 line monitoring: A comprehensive review. IEEE Access 9, 130410–130433
(2020) (2021)
24. Hao, Y., Liang, W., Yang, L., He, J., Wu, J.: Methods of image recognition 43. Jin, H., Lv, Z., Yuan, Z., Wei, Z., Wang, C., Wang, C., et al.: Micro-cracks
of overhead power line insulators and ice types based on deep weakly- identification and characterization on the sheds of composite insulators by
supervised and transfer learning. IET Gener. Transm. Distrib. 16(11), fractal dimension. IEEE Trans. Smart Grid 12(2), 1821–1824 (2021)
2140–2153 (2022) 44. Stefenon, S.F., Yow, K.C., Nied, A., Meyer, L.H.: Classification of distribu-
25. Qiu, Z., Zhu, X., Liao, C., Shi, D., Kuang, Y., Li, Y., et al.: Detection of tion power grid structures using inception v3 deep neural network. Electr.
bird species related to transmission line faults based on lightweight con- Eng. 104, 4557–4569 (2022)
volutional neural network. IET Gener. Transm. Distrib. 16(5), 869–881 45. Mussina, D., Irmanova, A., Jamwal, P.K., Bagheri, M.: Multi-modal data
(2022) fusion using deep neural network for condition monitoring of high voltage
26. Stefenon, S.F., Ribeiro, M.H.D.M., Nied, A., Mariani, V.C., Coelho, L.D.S., insulator. IEEE Access 8, 184486–184496 (2020)
Leithardt, V.R.Q., et al.: Hybrid wavelet stacking ensemble model for 46. Zhao, Z., Zhen, Z., Zhang, L., Qi, Y., Kong, Y., Zhang, K.: Insulator
insulators contamination forecasting. IEEE Access 9, 66387–66397 (2021) detection method in inspection image based on improved faster R-CNN.
27. Stefenon, S.F., Branco, N.W., Nied, A., Bertol, D.W., Finardi, E.C., Sartori, Energies 12(7), 1204 (2019)
A., et al.: Analysis of training techniques of ann for classification of insu- 47. Lei, X., Sui, Z.: Intelligent fault detection of high voltage line based on the
lators in electrical power systems. IET Gener. Transm. Distrib. 14(8), faster R-CNN. Measurement 138, 379–385 (2019)
1591–1597 (2020) 48. Guo, Y., Pang, Z., Du, J., Jiang, F., Hu, Q.: An improved AlexNet for power
28. Chen, K.L., Guo, Y., Ma, X.: Contactless voltage sensor for overhead edge transmission line anomaly detection. IEEE Access 8, 97830–97838
transmission lines. IET Gener. Transm. Distrib. 12(4), 957–966 (2018) (2020)
29. Xiong, S., Liu, Y., Yan, Y., Pei, L., Xu, P., Fu, X., et al.: Object recogni- 49. Sadykova, D., Pernebayeva, D., Bagheri, M., James, A.: IN-YOLO: Real-
tion for power equipment via human-level concept learning. IET Gener. time detection of outdoor high voltage insulators using UAV imaging.
Transm. Distrib. 15(10), 1578–1587 (2021) IEEE Trans. Power Delivery 35(3), 1599–1601 (2020)
30. He, T., Zeng, Y., Hu, Z.: Research of multi-rotor UAVs detailed 50. Rymarczyk, D., Struski, Ł., Górszczak, M., Lewandowska, K., Tabor, J.,
autonomous inspection technology of transmission lines based on route Zieliński, B.: Interpretable image classification with differentiable pro-
planning. IEEE Access 7, 114955–114965 (2019) totypes assignment. In: Computer Vision - ECCV 2022, pp. 351–368.
31. Waleed, D., Mukhopadhyay, S., Tariq, U., El-Hag, A.H.: Drone-based Springer Nature, Cham, Switzerland (2022)
ceramic insulators condition monitoring. IEEE Trans. Instrum. Meas. 70, 51. Yan, K., Li, Q., Li, H., Wang, H., Fang, Y., Xing, L., et al.: Deep learning-
1–12 (2021) based substation remote construction management and AI automatic
32. Salem, A.A., Abd-Rahman, R., Al Gailani, S.A., Salam, Z., Kamarudin, violation detection system. IET Gener. Transm. Distrib. 16(9), 1714–1726
M.S., Zainuddin, H., et al.: Risk assessment of polluted glass insulator using (2022)
leakage current index under different operating conditions. IEEE Access 52. Hsu, W.Y., Lin, W.Y.: Ratio-and-scale-aware YOLO for pedestrian detec-
8, 175827–175839 (2020) tion. IEEE Trans. Image Process. 30, 934–947 (2021)
STEFENON ET AL. 11

53. Souza, B.J., Stefenon, S.F., Singh, G., Freire, R.Z.: Hybrid-YOLO for clas- 62. Chen, C., Li, O., Tao, C., Barnett, A.J., Su, J., Rudin, C.: This looks like
sification of insulators defects in transmission lines based on UAV. Int. J. that: Deep learning for interpretable image recognition. arXiv 5, 1–12
Electr. Power Energy Syst. 148, 108982 (2023) (2019)
54. Tan, M., Pang, R., Le, Q.V.: Efficientdet: Scalable and efficient object detec- 63. Lewis, D., Kulkarni, P.: Insulator defect detection. IEEE Dataport (2021).
tion. In: Proceedings of the IEEE/CVF conference on computer vision https://www.doi.org/10.21227/vkdw-x769
and pattern recognition, pp. 10781–10790 (2020) 64. Serikbay, A., Bagheri, M., Zollanvari, A., Phung, B.T.: Accurate surface
55. Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., et al.: condition classification of high voltage insulators based on deep con-
Microsoft COCO: Common objects in context. In: European confer- volutional neural networks. IEEE Trans. Dielectr. Electr. Insul. 28(6),
ence on computer vision. Computer Vision, pp. 740–755. Springer, Cham, 2126–2133 (2021)
Switzerland (2014) 65. Alahyari, A., Hinneck, A., Tariverdizadeh, R., Pozo, D.: Segmentation and
56. Du, W., Shen, H., Fu, J.: Automatic defect segmentation in x-ray images defect classification of the power line insulators: A deep learning-based
based on deep learning. IEEE Trans. Ind. Electron. 68(12), 12912–12920 approach. In: 2020 International Conference on Smart Grids and Energy
(2021) Systems (SGES), pp. 476–481. IEEE, New York (2020)
57. Singh, G.: Think positive: An interpretable neural network for image 66. Zhang, Z., Huang, S., Li, Y., Li, H., Hao, H.: Image detection of insulator
recognition. Neural Networks 151, 178–189 (2022) defects based on morphological processing and deep learning. Energies
58. Stefenon, S.F., Singh, G., Yow, K.C., Cimatti, A.: Semi-ProtoPNet deep 15(7), 2465 (2022)
neural network for the classification of defective power grid distribution
structures. Sensors 22(13), 4859 (2022)
59. Singh, G., Yow, K.C.: Object or background: An interpretable deep learn-
ing model for COVID-19 detection from CT-scan images. Diagnostics How to cite this article: Stefenon, S.F., Singh, G.,
11(9), 1732 (2021) Souza, B.J., Freire, R.Z., Yow, K.-C.: Optimized hybrid
60. Singh, G., Yow, K.C.: An interpretable deep learning model for COVID-19 YOLOu-Quasi-ProtoPNet for insulators classification.
detection with chest x-ray images. IEEE Access 9, 85198–85208 (2021)
IET Gener. Transm. Distrib. 1–11 (2023).
61. Singh, G., Yow, K.C.: These do not look like those: An interpretable
deep learning model for image recognition. IEEE Access 9, 41482–41493 https://doi.org/10.1049/gtd2.12886
(2021)