ELYAN, E., JAMIESON, L. and ALI-GOMBE, A. 2020.

Deep learning for symbols detection and classification in

engineering drawings. Neural networks [online], 129, pages 91-102. Available from:
https://doi.org/10.1016/j.neunet.2020.05.025.

Deep learning for symbols detection and

classification in engineering drawings.

ELYAN, E., JAMIESON, L. and ALI-GOMBE, A.

2020

This document was downloaded from

https://openair.rgu.ac.uk
Deep Learning for Symbols Detection and Classification
in Engineering Drawings

Eyad Elyan∗, Laura Jamieson, Adamu Ali-Gombe

School of Computing Science and Digital Media, Robert Gordon University, UK

Abstract

Engineering drawings are commonly used in different industries such as Oil

and Gas, construction, and other types of engineering. Digitising these draw-
ings is becoming increasingly important. This is mainly due to the need to
improve business practices such as inventory, assets management, risk analy-
sis, and other types of applications. However, processing and analysing these
drawings is a challenging task. A typical diagram often contains a large number
of different types of symbols belonging to various classes and with very little
variation among them. Another key challenge is the class-imbalance problem,
where some types of symbols largely dominate the data while others are hardly
represented in the dataset. In this paper, we propose methods to handle these
two challenges. First, we propose an advanced bounding-box detection method
for localising and recognising symbols in engineering diagrams. Our method is
end-to-end with no user interaction. Thorough experiments on a large collec-
tion of diagrams from an industrial partner proved that our methods accurately
recognise more than 94% of the symbols. Secondly, we present a method based
on Deep Generative Adversarial Neural Network for handling class-imbalance.
The proposed GAN model proved to be capable of learning from a small num-
ber of training examples. Experiment results showed that the proposed method
greatly improved the classification of symbols in engineering drawings.
Keywords:

∗ Correspondingauthor
Email address: [email protected] (Eyad Elyan)
Deep Learning, YOLO, P&ID, Engineering Drawings, Symbols Recognition,
GANs

1. Introduction

Large volumes of un-digitised and paper-based documents are still very com-
mon across different domains. Amongst this legacy, engineering drawings are
known to be one of the most complex types of documents to process and anal-
5 yse. They are widely used in different industries such as construction and city
planning (i.e. Floor Plan diagrams [1]), Oil and Gas (i.e. P&IDs [2]), Me-
chanical Engineering [3], AutoCAD Drawing Exchange Format (DXF) [4] and
others. Interpreting these drawings requires highly skilled people, and in some
cases long hours of work.
10 In recent years, the digitisation of these drawings is becoming increasingly
important. This is partly due to the urgent need to improve business practices
such as inventory, assets management, risk analysis, safety checks and other
types of applications, and also due to the recent advancements in the domain of
machine vision and image understanding. Deep Learning (DL) [5], in particular,
15 had significantly improved the performance by orders of magnitude in many
domains such as Gaming and AI [6], Natural Language Processing [7], Health
[8], and others. One particular domain that has benefited hugely from DL is
machine vision [9]. Convolutional Neural Networks (CNNs) [10] have made
significant progress in recent years in many image-related tasks [11]. It has
20 been successfully applied to several fields such as hand-written recognition [12],
image classification [13, 14], Face Recognition & Biometrics [15] and others.
Before the CNNs, the improvements in image classification, segmentation, and
object detection was marginal and incremental. However, the introduction of
CNNs revolutionalised this field. For example, Deep Face [16], a face recognition
25 system that was first proposed by FaceBook in 2014 achieved an accuracy of
97.35%, beating the state-of-the-art then, by 27%.
Core image processing tasks such as shape and object detection, recognition,

2
and tracking have become much less challenging even under different conditions
and in much less controlled environments. Faster Region-based CNN (R-CNN)
30 [17], Single Shot Detectors (SSD) [18], Region-based Fully Convolutional Net-
works (R-FCN) [19] and You Only Look Once (YOLO) [20] are all relatively re-
cent methods that showed superior performance in the field of object detection,
tracking, and classification. These methods and their extensions have signifi-
cantly advanced this area of research and solved some of the most challenging
35 and inherent vision problems such as occlusions, light conditions, orientation,
and others, which were considered major challenges, even for a specific vision
task in a more controlled environment [21].
Significant advancement has also been made in the area of Generative Models
and was successfully applied in many applications. Among these, Generative
40 Adversarial Networks (GAN) proved to be one of the most established and
commonly used methods in generating content. GANs were initially introduced
by Ian Goodfellow in 2014 [22]. In the Methods section, we will discuss our
GAN-based method to handle the class imbalance problem. This is another
challenging problem that is common across many domains [23] including engi-
45 neering drawings, where one or more class of symbols in the diagrams are either
underrepresented or overrepresented in the dataset [24].
Despite this massive progress in the field of image processing and analysis,
very little progress has been made in the area of digitising complex engineering
drawings, and extracting information from these diagrams is still considered a
50 challenging problem [25]. To date, a major problem of most of the existing
solutions is that they still follow a traditional image-processing approach, which
requires extensive features extraction and engineering and carefully designed
heuristics [26]. These are often very domain-dependent, sensitive to noise and
data distribution, and mostly dedicated to solving part of the problem (i.e. de-
55 tecting symbols, separating graphics from text, and so on). As can be seen in
Figure 1, not only such an approach difficult to generalise across different sce-
narios, but also the performance of any machine learning algorithm will hugely
depend on the quality and accuracy of the extracted features.

3
Machine
Learning
Pre-process Features Algorithms

Input Features Colour Features

Images Extraction Shape Features
SIFT, SURF, ... Output/ Labels

Figure 1: Traditional frameworks for analysing images/ documents

In this paper, we propose an end-to-end framework for processing and analysing

60 complex engineering drawings. We argue that the core task of such a frame-
work is the accurate localisation and recognition of symbols in the drawing that
constitute a major part of it and simplifies subsequent tasks (i.e. line and text
detection). We show how one of the main inherent problems in classifying en-
gineering symbols, namely class-imbalance can be addressed using Generative
65 Adversarial Neural Networks. Figure 2 provides a schematic diagram of the
work presented in this paper. The main contributions of this work are outlined
as follows:

• We propose a novel pipeline for processing and analysing complex engi-

neering drawings. At the core of this pipeline is the accurate detection
70 and recognition of symbols.

• We show that an advanced-bounding-box detection method performs very

accurately on challenging engineering diagrams. To the best of our knowl-
edge, Deep Learning models (e.g. YOLO [27], RCNN [9] ) were never used
in such domain at a large scale of symbols with minimal difference. This is
75 mainly due to the complexity of the problem, and the very little variation
and noise within symbols of engineering drawings.

• Methods to handle the class-imbalance within engineering drawings are

presented and thoroughly evaluated. We present a fine-grained method to
train GAN models to generate engineering symbols of different overlapping
80 classes.

4
• Thorough evaluation using large collection of P&ID diagrams provided by
an industry partner in the Oil and Gas sector.

Deep Learning Augmented list of

Training recognised symbols in
Annotated Set YOLO-based methods the diagrams
P&ID
Diagrams

Testing
Set Model list of
recognised
symbols in MFC-GAN
the
diagrams

Figure 2: Schematic diagram of the framework for processing and analysing engineering doc-
uments

The rest of this paper is organised as follows: Section 2 presents an overview

and critical discussion of relevant work. In section 3, we present our methods,
85 dataset and pre-processing steps carried out. In section 4 we present our exper-
iments and discuss results. Finally, conclusions and future work are outlined in
section 5.

2. Related Work

In this section, we discuss relevant literature. First, we discuss literature

90 related to the processing and analysis of engineering drawings. This will be
followed by a brief introduction to Generative Adversarial Neural Networks and
how it can be applied to handle the class imbalance problem.

2.1. Engineering Drawings

An engineering drawing is a 2D image that contains different types of shapes,
95 symbols, lines, and text. These drawings are commonly used in different do-
mains and provide a rich representation of complex engineering workflows or
situations. The digitisation of these drawings was subject to extensive re-
search from the machine vision research community over the past four decades

5
[28, 29, 30, 31]. In recent years, and due to the significant progress in machine
100 vision research, computer power and also due to the availability of large vol-
umes of un-digitised data, the demand to have a fully automated framework for
digitising these drawings is becoming increasingly important.
Examples of work that aimed at extracting information from engineering
documents include analysis of musical notes [32], mechanical drawings [33], op-
105 tical character recognition (OCR) [34, 35, 36], and extracting information from
P&ID drawings [2, 37, 38]. It can be argued that most of the existing litera-
ture followed a traditional image processing approach [39], which requires some
form of feature extraction from the image[28], features representations [30], and
classification to determine the class of objects (i.e. symbols, digits, ...) [31].
110 The key limitation of traditional machine vision methods is that they re-
quire extensive features engineering, depend heavily on the quality of extracted
features, and often won’t generalise well to other unseen examples. A recent
extensive review showed that most of the existing literature focused on solving
part of the problem rather than providing a fully automated framework for digi-
115 tising an engineering diagram [26]. Examples include methods for recognising
symbols and lines in a drawing [40], detecting and separating text from symbols
and other graphics elements in diagram [38], classifying symbols in engineering
drawings [2] and so on. This is partly due to the complexity of the problem (i.e.
localising every single element in the document), and also due to the limitations
120 of the traditional image processing and analysis methods and the inherent vision
problems such as the sensitivity to noise, quality of the image, the orientation
of shapes and so on. Consider for example the work in [2], the authors used
a set of heuristic rules to localise symbols in the drawings, a Random Forest
[41] was then used to classify the symbols achieving an average accuracy higher
125 than 95%. Similar work was presented in [38], where a set of heuristics were
also used to detect and separate text from graphics elements. However, such an
approach is very dependent on the data distribution, and a slight variation in
the diagrams or in symbols representation might require adapting the existing
heuristics rules or creating new ones.

6
130 In a closely related area, Rebelo et al [42] presented a study on optical mu-
sic recognition and classification methods for musical symbols. They suggested
that adjoining staff lines, presence of symbols in close proximity to music notes,
broken symbols, overlapping symbols and areas with high symbol density all
contributed to the complexity of optical music recognition. Four classification
135 methods namely a multi-layer perceptron neural network model, Hidden Markov
Model, K-nearest neighbour and Support Vector Machine (SVM) were evalu-
ated on datasets of both synthetic and handwritten music scores. The highest
performance was obtained with an SVM model, however all approaches imple-
mented detection then removal of music staff lines and segmented the symbols
140 prior to symbol classification.
Khan et al [43], used video image analysis as part of a flight deck warning
system, which combined automated dial reading of flight instruments with do-
main knowledge. Experiments on a flight simulator and real flight aimed to
obtain the position of a white needle on the flight instrument using three image
145 processing approaches: background subtraction, pattern matching and a convo-
lution based approach. Results showed that the convolution method obtained
the highest accuracy, highest true positive rate and highest true negative rate.
In recent years, DL-based methods were explored and successfully applied
to some tasks that are similar to engineering drawings analysis. Ziran et al.
150 proposed a method, based on Single Shot Detectors (SSD) [18], to detect and
recognise furniture objects, doors, and windows in floor plan diagrams [44].
The results were encouraging. However, the datasets used were simple with a
limited number of furniture objects in each drawing (12). The performance also
dropped under the imbalanced class distribution of objects in the images.
155 Faster R-CNN was used in [45] for the detection and recognition of handwrit-
ten characters. Although the work focused mainly on specific elements of the
documents (mathematical expressions and flowcharts), promising results over
other traditional methods were achieved.
Detection and recognition of musical notes in documents have also benefited
160 from adopting Deep Learning-based methods [44, 46]. R-CNN, R-FCN, and

7
SDD were applied successfully to detect and recognise handwritten music notes
[46]. Results showed an improvement in symbols detection over other traditional
structured machine vision methods.
A framework for extracting information from P&ID drawings was presented
165 very recently in [47]. The authors used a two-step approach. First, Deep Learn-
ing methods were used to localise symbols and text, and then heuristic-based
methods were employed to detect other elements of the drawing (i.e. Euclidean
metrics for associating tags and symbols with pipelines, probabilistic Hough
transform to detect pipelines, etc.) The methods for localising symbols were
170 based on a fully connected convolutional neural network. A dataset of four
sheets consisting of 672 flow diagrams was used. Results were an improvement
over other traditional methods. However, accuracy wasn’t consistent across all
components. Class accuracy ranged from 100% for some components to 64.0%
for others (i.e. symbols of a certain class). Moreover, only a limited number of
175 symbols were used in this study (10 different classes of symbols) and the P&ID
sheets seem to be of a very good quality which is not often the case in the real
world.
To summarise, existing literature shows a clear gap between the current state
of machine vision and image understanding -due to the rapid development in this
180 field - and the slow and incremental progress in a very important application
domain across many industries.

2.2. GAN Models

Generative Adversarial Networks (GAN) were initially introduced by Ian

Goodfellow in 2014 [22]. These are considered as generative models that are
185 capable of producing new content. GANs are made of two contesting models
(i.e. Neural Networks, CNNs, etc), the Generator (G), and the Discriminator
(D). The discriminator is a classifier that receives input from the training set
(authentic content), and from the generator (fake input). During the training
process, the discriminator will learn how to distinguish between authentic and
190 fake input samples. On the other hand, the generator is trained to generate sam-

8
ples that capture the underlying characteristics of the original data (replicating
original content). Figure 3 depicts the GAN model.

Figure 3: Generative Adversarial Neural Networks

Adversarial training of both models G, and D is carried out using value

function as can be seen in Equation 1.

min max V (D, G) = Ex∼pdata (x) [logD(x)] + Ez∼pz (z) [log(1 − D(G(z)))] (1)
D G

195 Where pdata (x) is the probability distribution over the real data, x is a sam-
ple from the real training data, pz is the probability distribution over the noise
vector z, and G(z) is the output from the generator function G (or generated
images). GANs are state-of-the-art in terms of the quality of the image gener-
ated.
200 GANs have been successfully applied to different problems including image
generation [48, 49], segmentation and speech synthesis. In recent years they
were also successfully applied to handle class-imbalance problems [50, 51]. The
class imbalance is common across different domains including health, security,
and banking [23]. The problem happens when one or more class is either un-
205 derrepresented or overrepresented in the dataset. In such scenarios, a typical
supervised learning algorithm tends to be biased towards the majority class
when dealing with imbalanced datasets [24].
Supervised GANs provide an extension to the original GAN framework by
introducing conditional probabilities in the value function. This allows more
210 control over the generated samples and introduces diversity which is needed
for augmenting synthetic input data for class-imbalanced datasets. Typical

9
examples include vanilla GAN [22], CGAN [52] and AC-GAN [53]. Although
the literature shows that these models can be hugely affected by class-imbalance
especially in extreme cases [54].
215 Recent work appeared in [51] introduced a new extension of the GAN mod-
els. The authors trained the GAN models at a fine-grained level by updating the
discriminator objective to not only distinguish between fake and real instances
but also to classify the fake instances into different classes (i.e. Fake 1, Fake
2, etc). Extensive experiments using four different datasets showed superior re-
220 sults over other GAN models. Generated samples proved to be of good quality
and were successfully used to augment the dataset and improve the detection
rate of minority class instances.

3. Methods

Most engineering drawings contain a set of symbols, connectivity informa-

225 tion (lines) and some form of annotation (text). However, no public dataset is
available for evaluation purposes. In Section 3.1 we introduce our approach for
end-to-end symbols recognition from complex engineering drawings. The fol-
lowing subsection will discuss in detail the dataset used for experiments. This
will include data exploration and pre-processing. Finally, Section 3.4 provides
230 details of our proposed method to handle class-imbalance in these drawings.

3.1. Symbols Recognition

For locating and recognising symbols in the P&IDs, we propose to use

YOLO [55] method. This allows us to represent the problem as a set of bound-
ing box coordinates and class probabilities. The method is based on dividing
235 the entire image into S × S grid, where each cell predicts B bounding boxes and
confidence scores for those boxes [55]. The confidence scores are used to decide
if a cell contains a symbol or not. These are represented as a five-dimensional
vector (x, y, w,h, and confidence). Here, (x,y) represents the center of the
bounding box, while the width and height are predicted relative to the whole

10
240 image. The prediction from a grid is presented as S × S × (B ∗ 5 + C). Where
S is the size of the grid, B is the bounding box and C is the class probabilities
(i.e. probability of the symbol being gate valve, sensor, etc...). Figure 4 depicts
this setting.

Figure 4: The method divides the P&ID Diagram into a grid, following the YOLO model
[55], and predicts the class probabilities of the bounding boxes. The figure shows the symbols
sensor, flange, DBBPV, DB&BBV and RS

The YOLO model was chosen for two main reasons. First, it is a simple
245 framework, which allows simultaneous predictions of multiple bounding boxes
and class probabilities using a single convolutional neural network. Second, com-
pared to other models, YOLO is considered extremely fast. For testing P&IDs
that may contain on average 180 instances of various engineering symbols, this
is very important in a practical context.

250 3.2. Dataset - P&ID Diagrams

For experiments in this paper, we chose to work with Piping and Instru-
mentation Diagrams (P&IDs) Figure 5. A collection of 172 P&ID sheets were
obtained from an Oil and Gas industrial partner for evaluation purposes. These
diagrams contain different types of symbols, lines, and text (Figure 5).

11
Figure 5: Part of a P&ID Diagram

255 Additionally, the P&IDs are of different qualities, which makes the dataset
suitable for evaluation purposes. The P&ID diagrams can be defined as schematic
diagrams representing the different components of the process and the connectiv-
ity information. It is a representation of equipment (often depicted as symbols)
and process flow (depicted as different types of lines) [2].
260 Such diagrams are available across many industries in the form of paper or
scanned documents. Interpreting and analysing these documents requires expert
knowledge, and is often time-consuming [26, 56]. Moreover, a misinterpretation
of such documents can be very costly. For example, if a pipe needs to be replaced
in an Oil and Gas installation, then an engineer needs to check the corresponding
265 P&ID diagram, identify the valves that must be closed before carrying on the
task to ensure safety. In other words, accurate interpretation of these drawings
is paramount.

3.3. Data Exploration & Pre-processing

The original P&IDs sheets are large images, 7500 × 5250 pixels. To speed
270 up the training process we divided the sheet into 6 × 4 grid, resulting in 24 sub-

12
images (patches) with relatively much smaller sizes compared to the original
sheets (1250 × 1300).
Training a Deep Learning model requires fully annotated images/ diagrams.
1
To do so, we have used the Sloth tool to annotate the collection of P&ID
275 diagrams. In total 29 different symbols were annotated in the whole dataset
(Figure 7). The annotation process is simple and involves customising the sloth
tool to record the corresponding symbols names (class) and its location in the
diagram.
The resulting annotation of data is captured in a file representing the 29 sym-
280 bols. Data recorded included the x,y coordinates of the center of the bounding
boxes, width and height of the bounding box enclosing symbols. In total, 13,327
symbols belonging to the 29 different classes were annotated. The dataset is
hugely imbalanced as can be seen in Figure 6.

Figure 6: Class distribution of symbols in the whole dataset

Only 25 symbols of these were used in the experiments. These are shown
285 in Figure 7. Five symbols that were extremely under-represented in the whole

1 https://sloth.readthedocs.io/en/latest/

13
dataset (i.e. only one or two instances of these symbols appear in the training
and testing sets) were excluded from the first experiment.

Figure 7: Symbols used in the training and testing sets

3.4. MFC-GANs

To handle the class-imbalance in the dataset of engineering symbols (at the

290 classification level), we are proposing to use a method similar to the MFC-GAN
model presented in [51]. This model is chosen due to the very minor and in
some cases subtle difference between different classes of symbols. MFC-GAN
model allows us to train the discriminator to classify not only real symbols
but also fake symbols, which provides more fine-grained discrimination between
295 instances.
For this work, the discriminator network is designed to have four convolution
layers with strides of two and batch normalization is used between layers. All
convolution layers are activated using Leaky ReLu with alpha set to 0.2, and
Sigmoid function is used in the final layer as the activation function.
300 The discriminator layers are shared with a classifier model that outputs 2×N
soft-max. Where N is the number of classes. We also designed the generator to
have one linear layer and five transpose convolution layers with strides of two
in each layer. Batch normalization was also used between adjacent layers and
all layers were activated using Leaky ReLu apart from the final layer which is

14
305 sigmoid activated.
Similar to most GAN models the generator’s input is a noise vector of size
100 and combined with symbol label encoding (see [51] for details). This label
encoding is used to control the class-specific generation, which is essential for
our experiment.
310 The generator output is a 64 × 64 greyscale symbols image. For our ex-
periments (following sections) we used a batch size of 100 and a learning rate
of 0.001 which was experimentally chosen. Spectral normalisation was used in
both the generator and the discriminator. The proposed model will be trained
using Equations 2 and 3, 4.

Ls = E[log P (S = real|Xreal )] + E[log P (S = f ake|Xf ake )] (2)

Lcd = E[log P (C = c|Xreal )] + E[log P (C 0 = c0 |Xf ake )] (3)

Lcg = E[log P (C = c|Xreal )] + E[log P (C = c|Xf ake )] (4)

315 Where Ls is used to estimate the sampling loss, which represents the prob-
ability of the sample being real or fake. Lcd and Lcg are used to estimate the
classification losses over the generator and the discriminator. Xreal represents
the training data and Xf ake is the set of generated images.

4. Experiment & Results

320 Two experiments were carried out. The first experiment was designed to
evaluate an end-to-end solution for recognising symbols in engineering drawings.
We are assuming here, that locating and recognising these symbols will simplify
subsequent tasks in a framework for analysing the whole drawings (i.e. detecting
text, pipelines, etc...). This is simply because the majority of these types of
325 drawings are made of symbols. The second experiment is separate and is focused
on handling the class-imbalance problem using GAN-based methods.

15
4.1. Symbols Recognition

In our dataset, the P&ID sheets were approximately 7500 x 5250 pixels in
size. To use such image size in training data is computationally expensive and
330 therefore each P&ID was split into 24 patches by dividing the original P&ID
width by 6 and the height by 4. This gave a patch size of approximately 1250
x 1300 pixels. The annotation data for each patch was obtained using the
annotations for the whole P&ID as discussed in the previous section.
For the training phase, we excluded symbols that overlapped multiple patches.
335 After extensive experiments, it was decided to use the 3rd version of the YOLO
framework which proved to be improving the detection rate of small objects
compared to the first and second YOLO models [55], [27]. It is worth pointing
out that the sizes of the various engineering symbols in our dataset are relatively
small compared to the image size.
340 The YOLO architecture was customised for the purpose of this experiment.
First the number of classes in each of the three YOLO layers was set to 25, and
the number of filters was changed accordingly and was set to 3 × (Classno + 5),
where the Classno denotes the number of classes in the dataset.
The dataset was split approximately 90%:10% into training (155 P&IDs) and
345 test (16 P&IDs) sets. A pre-trained Network was used and retrained using our
dataset and all layers were fine-tuned. Darknet implementation of the YOLO
was used in this experiment 2 . During the training process the network input
size of 416 x 416 was adjusted after every ten batches; adjusting the input size
during training was reported to improve object detection across different object
350 scales [27]. The network was trained with a learning rate of 0.001 and training
is stopped when the model was trained on 10,000 batches, (batch size of 64).
At testing time, the model input was adjusted from 416×416 to 2400×2400.
In this way, we were able to test on the original P&ID images and simplify sym-
bol detection across a whole P&ID diagram in one step as opposed to combin-
355 ing detections from the P&ID patches. For evaluation, symbols were compared

2 https://github.com/AlexeyAB/darknet, A. A.B., Darknet,(2019)

16
against the ground truth and the Intersection over Union IOU was set experi-
mentally to 0.5. A simple front end was developed using Python Libraries and
OpenCV3 for visualisation and manual error analysis purposes.

4.1.1. Results
360 The training accuracy achieved was ∼ 96%. On the testing set, 1352 sym-
bols out of 1424 were correctly located and recognised with a testing accuracy
equal to ∼ 94.9%. A heatmap of the confusion matrix for the testing set is pre-
sented in Figure 8. It can be seen and as expected that majority class instances
were accurately detected and recognised. In other words, symbols with enough
365 examples in the training set were accurately recognised.

Figure 8: Heatmap of the Confusion matrix of the 25 symbols predictions (Testing Set)

3 https://opencv.org/

17
A typical output from the proposed methods where different symbols are high-
lighted in different colours is shown Figure 9 . Recognised symbols here were
numbered and the predicted labels were recorded for further comparison against
the ground truth. These symbols include inlets/outlets that are denoted by la-
370 bel to, label from, sensors, ball valve, reducers, gate valves, globe valves, and
others.

Figure 9: A P&ID diagram with various recognised symbols (Testing Set)

18
Table 1 provides more details about the number of instances of each symbol
in the training and testing set named as No of Training Symbols and No of
Testing Symbols respectively. Furthermore, it also shows the number of correctly
375 recognised symbols in the testing set (Correctly Recognised) and the testing
accuracy per class (Class Accuracy).

Table 1: Results of the proposed methods for symbols recognition (Testing Set)

Symbol No of Training Symbols No of Testing Symbols Correctly Recognised (Testing) Class Accuracy
Sensor 2810 302 297 98%
Ball Valve 1629 213 212 99%
Label From 1347 103 103 100%
Label To 1178 113 113 100%
Flange 1110 158 121 77%
reducer 821 91 90 99%
DB&BBV 542 67 66 98%
Gate Valve 535 110 104 94%
check valve 396 42 42 100%
TOB/Butterfly Valve 178 59 58 98%
Plug Valve 173 8 8 100%
Globe Valve 161 7 7 100%
Needle Valve 160 10 10 100 %
RS 143 26 24 92%
PSV 118 25 22 88%
eccentric reducer 98 23 22 96%
POB valve 84 16 16 100%
DBBPV 83 15 15 100%
PRV 32 8 8 100%
control valve globe 30 6 6 100%
control valve 22 5 5 100%
vent to atm 19 8 2 25%
injection/sample point 13 2 1 50%
Angle Valve 11 2 0 0.0%
BPRV 11 5 0 0.0%

Results show that the majority of instances were accurately detected and
recognised (1352 of 1424). Figure 10 shows different symbols from various P&ID
diagrams. Notice here that symbols are accurately detected and recognised
380 regardless of its orientation. For example, reducers, gate valves, check valves,
and others appear in different orientations (Figure 10). Similarly, sensors are
accurately detected and recognised regardless of the text overlap with these
instances. This clearly shows that unlike traditional methods the proposed

19
method is robust to these inherent vision challenges (at least in this context).

Figure 10: Examples of detected symbols

385 As can be seen in Table 2, in total 72 instances of the P&ID symbols were
either unrecognised at all (missed), or incorrectly classified as different symbols.
Of these, 8 instances of symbols were incorrectly classified (Table 2). Addi-
tionally, 64 symbols were completely missed. This can be largely attributed
to the nature of the drawings, wherein these cases symbols will have text and
390 annotation almost covering its entirety. This is evident if we look at the IOU
in Table 2 which is zero across all these missed symbols.

20
Table 2: Unrecognised and misclassified symbols in Engineering Drawings

Actual Class No of Instances Predicted Class IOU

Ball Valve 1 reducer 0.81
BPRV 5 PRV 0.91
eccentric reducer 1 reducer 0.72
reducer 1 eccentric reducer 0.90
Angle Valve 2 - 0.00
Flange 37 - 0.00
Gate Valve 6 - 0.00
Sensor 5 - 0.00
TOB/Butterfly Valve 1 - 0.00
Vent to Atm 6 - 0.00
injection/sample point 1 - 0.00
PSV 3 - 0.00
RS 2 - 0.00
DB&BBV 1 - 0.00

Further visual analysis of the results presented in Table 2 showed that some
symbols were incorrectly labeled. In particular, the instance of the symbol
Ball Valve, although the model predicted the ’wrong’ class symbol, visualising
395 the results showed that the model actually predicted the right class for these
symbols despite the wrong label. This is illustrated in Figure 11. Here, we use
a simple front end to visualise the recognised symbols alongside an item number
that we assign for each of them. This has greatly facilitated the analysis and
visualisation of the results.

21
Figure 11: Incorrectly labeled symbol

400 Similarly, consider the symbol of class BPRV which was classified as PRV
in all five instances in the testing set. First, it is worth noting that the number
of training instances of this symbol is extremely low (11). Additionally, the
symbol is very much similar to the PRV class. However, it is anticipated that
more training examples of this symbol will certainly improve its detection rate,
405 as it is the case with most majority class symbols (i.e. Sensor, Ball Valve,
Reducer, Gate Valve, Check Valve, Globe Valve and so on). Figure 12 shows
samples of the BPRV symbols which were incorrectly classified alongside the
actual PRV symbol.

Figure 12: Incorrectly classified BPRV symbols (first four instances from left) as PRV symbol
(fifth instance)

In summary, it can also be seen from the results presented in Table 1 that

22
410 the recognition rate of the symbols (vent to atm, Angle Valve, and BPRV ) was
quite low. This is mainly due to the limited numbers of training samples that
represent these symbols. But overall, and excluding these three symbols, the
average class accuracy of the remaining 22 symbols in the dataset is over 92%
which is very encouraging results for such a challenging problem.

415 4.2. MFC-GAN for Class-Imbalance

In this experiment we aim at evaluating a GAN-based model to handle the

class-imbalance problem in the dataset. This is not a recognition task as in
the first experiment but rather a classification problem. The experiment aims
first at generating more symbols using MFC-GAN model. Then these synthe-
420 sized samples will be used to augment the training set aiming at improving
classification results.
The dataset used in this experiment is the almost the same one used in
Experiment 1. All symbols were resized to 64 × 64 grey-scaled images. The
problem is formulated as a supervised learning task where the aim is to learn a
425 function f (x) that maps an instance xi of a particular engineering symbol to the
corresponding class yi . In this case, yi ∈ Y where Y is a discrete set of classes
representing the 29 symbols in the dataset. As discussed earlier the dataset is
hugely imbalanced, and some of the instance that were dropped in Experiment
1 populates less than ∼ 0.01% such as angle choke valve.
430 The experiment was carried out in two stages, a GAN training stage and
a classification stage. First, we trained MFC-GAN using all the samples in
the dataset. The MFC-GAN model was conditioned to generate engineering
symbols in extreme cases of class imbalance. To do so, we considered the least
represented symbols in the whole dataset. These are Angle Choke Valve, Angle
435 Valve, Injection Sample Point, Back Pressure Regulating Valve, PS Gate Valve,
Control Valve, Through Conduit Gate Valve, Control Valve Globe and Pressure
Regulating Valve. These symbols have 2, 13, 15, 17, 17, 27, 31, 36 and 42
instances respectively in the training set. The model was trained only once
on this dataset and the samples were generated after training was completed.

23
440 During training, the minority classes were resampled to encourage learning of
minority instances structure.
The trained MFC-GAN model was then used to generate symbols of minority
class instances (the least represented in the dataset, nine symbols). The original
dataset was split 70% for training and 30% for testing set. Synthetic datasets
445 were then added to the training set. For each minority class, 5000 more synthetic
samples were added. This enabled us to rebalance the dataset by increasing the
presence of the least represented symbols.
In order to evaluate the quality of the generated symbols, we build a classi-
fication model to compare performance before and after adding the generated
450 symbols to the training set.
The classification model chosen is a CNN with 4 layers. The first three
layers are convolution layers with 32, 64, 128 outputs. These layers have a
kernel size of 3 × 3, 2 × 2 and max-pooling in-between them. The fourth layer is
a fully connected layer with 256 units that feeds in to a 29-way SoftMax output
455 representing the 29 symbol classes. The CNN was trained using SGD with a
batch size of 64 and a learning rate of 0.001. Classification results were recorded
using common metrics, namely true positive rate, balanced accuracy, G-mean
and F1-Score.

4.2.1. Results
460 Figure 13 compares the generated samples from MFC-GAN model with the
original symbols from the diagram. We also report the symbols classification
results in Table 3.
MFC-GAN generated far superior and more realistic samples. Visual in-
spection revealed distinct symbols features and the required categories were
465 generated in each instance. Moreover, MFC-GAN high-quality samples had a
positive effect on the performance of the classifier. For example, the G-Mean
and sensitivity improved from 0 to 100% on angle choke valve as can be seen
in Table 3 with just two instances of the classes. This result is consistent in
seven of the nine minority classes. However, we observed that the model did

24
P&ID MFC-GAN sample

Figure 13: Comparing original P&ID samples with MFC-GAN generated samples.

470 not improve the baseline in the other two classes control valve and PRV classes.
A closer look at Figure 13 revealed a high similarity between symbols. There is
extreme similarity between angle valve (fifth symbol from the top) with control
valve globe (eighth symbols from the top) and PRV (seventh symbol from the
top) and BPRV (second symbol from the top). Although symbols were dis-
475 tinctly generated, the similarity of symbols dwindled the classification results in
these classes. The low precision in BPRV and control valve globe classes from
Table 3 further solidifies this observation.

Table 3: CNN performance on symbols classification.

Metric Model angle choke valve Angle Valve BPRV control valve control valve globe injectionsample point PRV PS Gate Valve TCGvalve

Baseline 0.00 0.50 0.60 0.88 1.00 0.80 1.00 1.00 0.89
Sensitivity
MFC-GAN 1.00 1.00 0.80 0.88 1.00 0.88 0.77 1.00 0.91

Baseline 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
Specificity
MFC-GAN 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00

Baseline 0.00 1.00 1.00 1.00 0.85 1.00 0.72 1.00 1.00
Precision
MFC-GAN 1.00 1.00 0.67 1.00 0.92 1.00 0.91 0.83 1.00

Baseline 0.00 0.67 0.75 0.93 0.92 0.89 0.84 1.00 0.94
F1-score
MFC-GAN 1.00 1.00 0.73 0.93 0.96 0.93 0.83 0.91 0.95

Baseline 0.50 0.75 0.80 0.94 1.00 0.90 1.00 1.00 0.95
Accuracy
MFC-GAN 1.00 1.00 0.90 0.94 1.00 0.94 0.89 1.00 0.96

Baseline 0.00 0.71 0.77 0.94 1.00 0.89 1.00 1.00 0.94
G-Mean
MFC-GAN 1.00 1.00 0.89 0.94 1.00 0.93 0.88 1.00 0.95

MFC-GAN models proved in this experiment to be able to generate minor-

ity class instances that are extremely under-represented in the dataset. The

25
480 quality of these samples was evaluated subjectively by inspecting the resulting
samples, and objectively by measuring a classifier performance before and after
adding the generated samples to the training sets. Results show clearly that
performance improved across several common evaluation metrics. However, it
has to be said that MFC-GAN is only one method that can be used to han-
485 dle the class imbalance problem. Other possible methods can also be explored
and utilized. Class-imbalance is a very well researched problem, and there is a
wide range of methods that ranges from simple data augmentation, sampling to
more advanced methods such as GAN [24]. For an extensive review of different
possible methods, the reader is referred to [57].

490 5. Conclusion & Future Direction

In this paper, we proposed an end-to-end framework for processing and

analysing complex engineering drawings. Thorough experiments using a large
collection of P&ID sheets from an industrial partner showed that our method
accurately recognises more than 94% of the symbols in the drawings. Advanced-
495 bounding-box detection methods proved in our experiments that they perform
accurately in such challenging tasks by recognising symbols of 25 different
classes, despite the very little differences between some of these symbols. Ad-
ditionally, we proposed a GAN-based model to handle class-imbalance in the
symbols dataset. Our experiments demonstrated that our method was capable
500 of generating plausible engineering symbols and also proved to be improving
classification accuracy when augmenting the training set with this synthesized
data. Experiments results show that the proposed GAN model can learn from
a smaller number of training examples.
A future direction of this work will focus on utilising Generative Adversarial
505 Neural Networks to generating symbols in a diagram context. In other words,
generate part of the engineering diagram, and not only the symbols. This will
greatly help in saving efforts needed for manual data annotation. Additionally,
future work will include building a unified framework based on the proposed

26
methods to allow full processing and analysis of engineering diagrams such as
510 P&ID. We hypothesize that the work presented in this paper will greatly simplify
subsequent tasks such as text localisation and line detection.

Acknowledgement

This work was supported by Scotland Data Lab Innovation Centre, Oil and
Gas Innovation Centre and DNV GL.

515 References

[1] S. Ahmed, M. Liwicki, M. Weber, A. Dengel, Automatic room detection

and room labeling from architectural floor plans, in: 2012 10th IAPR In-
ternational Workshop on Document Analysis Systems, 2012, pp. 339–343.
doi:10.1109/DAS.2012.22.

520 [2] E. Elyan, C. M. Garcia, C. Jayne, Symbols classification in engineering

drawings, in: 2018 International Joint Conference on Neural Networks
(IJCNN), 2018, pp. 1–8. doi:10.1109/IJCNN.2018.8489087.

[3] P. Vaxiviere, K. Tombre, Celesstin: Cad conversion of mechanical drawings,

Computer 25 (7) (1992) 46–54. doi:10.1109/2.144439.

525 [4] K. N. Goh, S. R. Mohd. Shukri, R. B. H. Manao, Automatic assessment for

engineering drawing, in: H. B. Zaman, P. Robinson, P. Olivier, T. K. Shih,
S. Velastin (Eds.), Advances in Visual Informatics, Springer International
Publishing, Cham, 2013, pp. 497–507.

[5] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT Press, 2016,

530 http://www.deeplearningbook.org.

[6] S. D. Holcomb, W. K. Porter, S. V. Ault, G. Mao, J. Wang, Overview on

deepmind and its alphago zero ai, in: Proceedings of the 2018 International
Conference on Big Data and Education, ICBDE ’18, ACM, New York, NY,

27
USA, 2018, pp. 67–71. doi:10.1145/3206157.3206174.
535 URL http://doi.acm.org/10.1145/3206157.3206174

[7] Z. Yang, D. Yang, C. Dyer, X. He, A. Smola, E. Hovy, Hierarchical at-

tention networks for document classification, in: Proceedings of the 2016
Conference of the North American Chapter of the Association for Com-
putational Linguistics: Human Language Technologies, Association for
540 Computational Linguistics, San Diego, California, 2016, pp. 1480–1489.
doi:10.18653/v1/N16-1174.
URL https://www.aclweb.org/anthology/N16-1174

[8] A. Esteva, A. Robicquet, B. Ramsundar, V. Kuleshov, M. DePristo,

K. Chou, C. Cui, G. Corrado, S. Thrun, J. Dean, A guide to deep learn-
545 ing in healthcare, Nature Medicine 25 (1) (2019) 24–29. doi:10.1038/
s41591-018-0316-z.
URL https://doi.org/10.1038/s41591-018-0316-z

[9] R. Girshick, Fast r-cnn, in: Proceedings of the 2015 IEEE International
Conference on Computer Vision (ICCV), ICCV ’15, IEEE Computer So-
550 ciety, Washington, DC, USA, 2015, pp. 1440–1448. doi:10.1109/ICCV.
2015.169.
URL http://dx.doi.org/10.1109/ICCV.2015.169

[10] J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu,

X. Wang, G. Wang, J. Cai, T. Chen, Recent advances in convolu-
555 tional neural networks, Pattern Recognition 77 (2018) 354 – 377.
doi:https://doi.org/10.1016/j.patcog.2017.10.013.
URL http://www.sciencedirect.com/science/article/pii/
S0031320317304120

[11] A. Ali-Gombe, E. Elyan, C. Jayne, Fish classification in context of noisy

560 images, in: G. Boracchi, L. Iliadis, C. Jayne, A. Likas (Eds.), Engineering
Applications of Neural Networks, Springer International Publishing, Cham,
2017, pp. 216–226.

28
[12] Y. Lecun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning ap-
plied to document recognition, Proceedings of the IEEE 86 (11) (1998)
565 2278–2324. doi:10.1109/5.726791.

[13] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,

V. Vanhoucke, A. Rabinovich, Going deeper with convolutions, in: 2015
IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
2015, pp. 1–9. doi:10.1109/CVPR.2015.7298594.

570 [14] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with

deep convolutional neural networks, Commun. ACM 60 (6) (2017) 84–90.
doi:10.1145/3065386.

[15] U. Park, A. K. Jain, Face matching and retrieval using soft biometrics,
IEEE Transactions on Information Forensics and Security 5 (3) (2010) 406–
575 415. doi:10.1109/TIFS.2010.2049842.

[16] Y. Taigman, M. Yang, M. Ranzato, L. Wolf, Deepface: Closing the gap

to human-level performance in face verification, in: 2014 IEEE Conference
on Computer Vision and Pattern Recognition, 2014, pp. 1701–1708. doi:
10.1109/CVPR.2014.220.

580 [17] S. Ren, K. He, R. Girshick, J. Sun, Faster r-cnn: Towards real-time object
detection with region proposal networks, in: Proceedings of the 28th Inter-
national Conference on Neural Information Processing Systems - Volume
1, NIPS’15, MIT Press, Cambridge, MA, USA, 2015, pp. 91–99.
URL http://dl.acm.org/citation.cfm?id=2969239.2969250

585 [18] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. E. Reed, C. Fu, A. C.

Berg, SSD: single shot multibox detector, CoRR abs/1512.02325. arXiv:
1512.02325.
URL http://arxiv.org/abs/1512.02325

[19] J. Dai, Y. Li, K. He, J. Sun, R-FCN: object detection via region-based fully

29
590 convolutional networks, CoRR abs/1605.06409. arXiv:1605.06409.
URL http://arxiv.org/abs/1605.06409

[20] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Uni-
fied, real-time object detection, in: Proceedings of the IEEE conference on
computer vision and pattern recognition, 2016, pp. 779–788.

595 [21] W. Zhao, R. Chellappa, P. J. Phillips, A. Rosenfeld, Face recognition: A

literature survey, ACM Comput. Surv. 35 (4) (2003) 399–458. doi:10.
1145/954339.954342.
URL http://doi.acm.org/10.1145/954339.954342

[22] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,

600 S. Ozair, A. Courville, Y. Bengio, Generative adversarial nets, in: Ad-
vances in neural information processing systems, 2014, pp. 2672–2680.

[23] P. Vuttipittayamongkol, E. Elyan, A. Petrovski, C. Jayne, Overlap-based

undersampling for improving imbalanced data classification, in: H. Yin,
D. Camacho, P. Novais, A. J. Tallón-Ballesteros (Eds.), Intelligent Data
605 Engineering and Automated Learning – IDEAL 2018, Springer Interna-
tional Publishing, Cham, 2018, pp. 689–697.

[24] P. Vuttipittayamongkol, E. Elyan, Neighbourhood-based undersampling

approach for handling imbalanced and overlapped data, Information
Sciences 509 (2020) 47 – 70. doi:https://doi.org/10.1016/j.ins.
610 2019.08.062.
URL http://www.sciencedirect.com/science/article/pii/
S0020025519308114

[25] E. Arroyo, A. Fay, M. Chioua, M. Hoernicke, Integrating plant and process

information as a basis for automated plant diagnosis tasks, in: Proceedings
615 of the 2014 IEEE Emerging Technology and Factory Automation (ETFA),
2014, pp. 1–8. doi:10.1109/ETFA.2014.7005098.

30
[26] C. F. Moreno-Garcı́a, E. Elyan, C. Jayne, New trends on digitisation of
complex engineering drawings, Neural Computing and Applicationsdoi:
10.1007/s00521-018-3583-1.
620 URL https://doi.org/10.1007/s00521-018-3583-1

[27] J. Redmon, A. Farhadi, Yolo9000: Better, faster, stronger, in: 2017 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), 2017,
pp. 6517–6525. doi:10.1109/CVPR.2017.690.

[28] A. K. Chhabra, Graphics Recognition Algorithms and Systems, in: Pro-

625 ceedings of the 2nd International Conference on Graphics Recognition
(GREC’97 ), 1997, pp. 244–252. doi:10.1007/3-540-64381-8_40.

[29] L. P. Cordella, M. Vento, Symbol recognition in documents: A collection of

techniques?, International Journal on Document Analysis and Recognition
3 (2) (2000) 73–88. doi:10.1007/s100320000036.

630 [30] D. Zhang, G. Lu, Review of shape representation and description tech-
niques, Pattern Recognition 37 (1) (2004) 1–19. doi:10.1016/j.patcog.
2003.07.008.

[31] S. V. Ablameyko, S. Uchida, Recognition of engineering drawing entities:

Review of approaches, International Journal of Image and Graphics 07 (04)
635 (2007) 709–733. arXiv:http://www.worldscientific.com/doi/pdf/10.
1142/S0219467807002878, doi:10.1142/S0219467807002878.

[32] D. Blostein, General Diagram-Recognition Methodologies, in: Proceedings

of the 1st International Conference on Graphics Recognition (GREC’95),
1995, pp. 200–212.

640 [33] T. Kanungo, R. M. Haralick, D. Dori, Understanding Engineering Draw-

ings: A Survey, in: Proceedings of the 1st International Conference on
Graphics Recognition (GREC’95), 1995, pp. 119–130.

31
[34] C. R. Kulkarni, A. B. Barbadekar, Text Detection and Recognition: A Re-
view, International Research Journal of Engineering and Technology (IR-
645 JET) 4 (6) (2017) 179–185.

[35] Y. Lu, Machine printed character segmentation - An overview, Pattern

Recognition 28 (1) (1995) 67–80. doi:10.1016/0031-3203(94)00068-W.

[36] S. Mori, C. Y. Suen, K. Yamamoto, Historical Review of OCR Research

and Development, Proceedings of the IEEE 80 (7) (1992) 1029–1058. doi:
650 10.1109/5.156468.

[37] C. Howie, J. Kunz, T. Binford, T. Chen, K. Law, Computer in-

terpretation of process and instrumentation drawings, Advances
in Engineering Software 29 (7) (1998) 563 – 570. doi:https:
//doi.org/10.1016/S0965-9978(98)00022-2.
655 URL http://www.sciencedirect.com/science/article/pii/
S0965997898000222

[38] C. F. Moreno-Garcı́a, E. Elyan, C. Jayne, Heuristics-Based Detection to

Improve Text / Graphics Segmentation in Complex Engineering Drawings,
in: Engineering Applications of Neural Networks, Vol. CCIS 744, 2017, pp.
660 87–98.

[39] R. C. Gonzalez, R. E. Woods, Digital image processing, Prentice Hall,

Upper Saddle River, N.J., 2008.
URL http://www.amazon.com/Digital-Image-Processing-3rd-Edition/
dp/013168728X

665 [40] L. Boatto, V. Consorti, M. Del Buono, V. Eramo, A. Esposito, F. Melcarne,

M. Meucci, A. Morelli, M. Mosciatti, A. Spirito, M. Tucci, Detection and
separation of symbols connected to graphics in line drawings, in: Proceed-
ings., 11th IAPR International Conference on Pattern Recognition. Vol.II.
Conference B: Pattern Recognition Methodology and Systems, 1992, pp.
670 545–548. doi:10.1109/ICPR.1992.201837.

32
[41] E. Elyan and M. M. Gaber, A genetic algorithm approach to optimising
random forests applied to class engineered data, Information Sciences 384
(2017) 220 – 234. doi:https://doi.org/10.1016/j.ins.2016.08.007.
URL http://www.sciencedirect.com/science/article/pii/
675 S0020025516305783

[42] A. Rebelo, G. Capela, J. S. Cardoso, Optical recognition of music symbols:

A comparative study, Int. J. Doc. Anal. Recognit. 13 (1) (2010) 1931.
doi:10.1007/s10032-009-0100-1.
URL https://doi.org/10.1007/s10032-009-0100-1

680 [43] W. Khan, D. Ansell, K. Kuru, M. Bilal, Flight guardian: Autonomous flight
safety improvement by monitoring aircraft cockpit instruments, Journal
of Aerospace Information Systems 15 (4) (2018) 203–214. arXiv:https:
//doi.org/10.2514/1.I010570, doi:10.2514/1.I010570.
URL https://doi.org/10.2514/1.I010570

685 [44] A. Pacha, J. Haji, J. Calvo-Zaragoza, A baseline for general music ob-
ject detection with deep learning, Applied Sciences 8 (9). doi:10.3390/
app8091488.
URL http://www.mdpi.com/2076-3417/8/9/1488

[45] F. D. Julca-Aguilar, N. S. T. Hirata, Symbol detection in online hand-

690 written graphics using faster R-CNN, CoRR abs/1712.04833. arXiv:
1712.04833.
URL http://arxiv.org/abs/1712.04833

[46] A. Pacha, K. Choi, B. Coasnon, Y. Ricquebourg, R. Zanibbi, H. Eiden-

berger, Handwritten music object detection: Open issues and baseline re-
695 sults, in: 2018 13th IAPR International Workshop on Document Analysis
Systems (DAS), 2018, pp. 163–168. doi:10.1109/DAS.2018.51.

[47] R. Rahul, S. Paliwal, M. Sharma, L. Vig, Automatic information extraction

from piping and instrumentation diagrams, CoRR abs/1901.11383. arXiv:

33
1901.11383.
700 URL http://arxiv.org/abs/1901.11383

[48] A. Ali-Gombe, E. Elyan, Y. Savoye, C. Jayne, Few-shot classifier gan, in:

2018 International Joint Conference on Neural Networks (IJCNN), 2018,
pp. 1–8. doi:10.1109/IJCNN.2018.8489387.

[49] A. Ali-Gombe, E. Elyan, C. Jayne, Multiple fake classes gan for data aug-
705 mentation in face image dataset, in: 2019 International Joint Conference
on Neural Networks (IJCNN), 2019, pp. 1–8. doi:10.1109/IJCNN.2019.
8851953.

[50] A. Antoniou, A. Storkey, H. Edwards, Data augmentation generative ad-

versarial networks, arXiv preprint arXiv:1711.04340.

710 [51] A. Ali-Gombe, E. Elyan, Mfc-gan: Class-imbalanced dataset classification

using multiple fake class generative adversarial network, Neurocomputing
361 (2019) 212–221.

[52] M. Mirza, S. Osindero, Conditional generative adversarial nets, arXiv

preprint arXiv:1411.1784.

715 [53] A. Odena, C. Olah, J. Shlens, Conditional image synthesis with auxiliary
classifier gans, International conference on machine learning,page 2642-2651
70 (AUG 2017) 2642–2651.

[54] G. Mariani, F. Scheidegger, R. Istrate, C. Bekas, C. Malossi, Bagan: Data

augmentation with balancing gan, arXiv preprint arXiv:1803.09655.

720 [55] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once:
Unified, real-time object detection, in: 2016 IEEE Conference on Com-
puter Vision and Pattern Recognition (CVPR), 2016, pp. 779–788. doi:
10.1109/CVPR.2016.91.

[56] E. Arroyo, X. L. Hoang, A. Fay, Automatic detection and recognition of

725 structural and connectivity objects in svg-coded engineering documents, in:

34
2015 IEEE 20th Conference on Emerging Technologies Factory Automation
(ETFA), 2015, pp. 1–8. doi:10.1109/ETFA.2015.7301510.

[57] Learning from class-imbalanced data: Review of methods and applications,

Expert Systems with Applications 73 (2017) 220 – 239.

35