Private Body Part Detection using Deep Learning

André Tabone

1 a

, Alexandra Bonnici

1 b

, Stefania Cristina

1 c

, Reuben Farrugia

2 d

and Kenneth Camilleri

1 e

Department of Systems and Control, University of Malta, Malta

Department of Computer Engineering, University of Malta, Malta

Keywords:

Deep Neural Networks, Pornographic Detection, Classiﬁers, Private Body Part Detector.

Abstract:

Fast and accurate detection of sexually exploitative imagery is necessary for law enforcement agencies to allow

for prosecution of suspect individuals. In literature, techniques which can be used to assist law enforcement

agencies only determine whether the image content is pornographic or benign. In this paper, we provide a

review on classical handcrafted-feature based and deep-learning based pornographic detection in images and

describe a framework which goes beyond this, to identify the location of genitalia in the image. Despite this

being a computationally complex task, we show that by learning multiple features, a MobileNet framework

can achieve an accuracy of 76.29% in the correct labelling of female and male sexual organs.

1 INTRODUCTION

Every month, a new mass of data pertaining to

pornography is made available on the Internet (Vi-

torino et al., 2018). Tracking such data to curtail

the sharing of pornography and prosecute perpetra-

tors is a critical aspect of cyber-crime activities. Soft-

ware capable of fast detection of pornographic con-

tent is essential to law enforcement agencies (LEAs)

and such needs have driven researchers to propose al-

gorithms which aid LEAs in their ﬁght against cy-

bercrime. Thus, tools that detect pornographic con-

tent (Wehrmann et al., 2018), perform age estima-

tion (Macedo et al., 2018b), and search for speciﬁc

keywords in ﬁle names, amongst others, exist to aid

LEAs. In the analysis of pornographic image content,

algorithms which determine whether images contain

exposed private body parts provide LEAs with essen-

tial descriptors of the image content. Such descriptors

provide a better understanding of the reason why an

images is pornographic without the need for manual

inspection.

Deep learning algorithms are proving to be very

effective in object detection and classiﬁcation. To our

https://orcid.org/0000-0002-7144-5221

https://orcid.org/0000-0002-6580-3424

https://orcid.org/0000-0003-4617-7998

https://orcid.org/0000-0001-8106-9891

https://orcid.org/0000-0003-0436-6408

knowledge, however, deep learning algorithms which

analyse images speciﬁcally to determine whether

these contain private body parts do not exist. The po-

sition taken by this paper is that by training a deep

network on windows containing private body parts,

we can create a system which detects and labels these

parts. To this extent, we propose a two-step approach,

in which the ﬁrst step is a pornographic image detec-

tor, and the second step uses a windowing approach

to detect private body parts within the image.

The rest of the paper is organised as follows. Sec-

tion 2 gives an overview of the related work in the

area, Section 3 describes our proposed approach, Sec-

tion 4 presents the results obtained, while Section 5

concludes the paper.

2 LITERATURE REVIEW

Human skin offers a practical feature for pornogra-

phy detection since it is invariant to partial occlusion,

as well as to changes in scaling and rotation (Naji

et al., 2018). However, skin detection is an insufﬁ-

cient indicator of pornography since instances of be-

nign images may also have large areas of exposed

skin (Wang et al., 2009). Thus, skin detection of-

ten acts as a precursor to sexual organ detection, par-

ticularly, of the female breast (Fuangkhon and Tan-

prasert, 2005). In such cases, skin detection is fol-

lowed by either model-based detection which uses ge-

Tabone, A., Bonnici, A., Cristina, S., Farrugia, R. and Camilleri, K.

Private Body Part Detection using Deep Learning.

DOI: 10.5220/0009101502050211

In Proceedings of the 9th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2020), pages 205-211

ISBN: 978-989-758-397-1; ISSN: 2184-4313

205

Table 1: Method and performance comparison of algorithms in previous literature.

Method Aim Approach Performance

(Shen et al., 2010)

Sexual organ detection Human pose model True positive rate: 89%, False positive rate: 10%

(Choi et al., 2011)

Pornographic classiﬁcation

Skin detection + Colour and

texture features + MAP

True positive rate: 93.63%, False positive rate: 10.13%

False negative rate: 6.9%

(Tian et al., 2018)

Sexual organ detection Deformable Part Model + SVM Overall precision: 80%, Recall: 82%, F-1 score: 18%

(Lv et al., 2011)

Pornographic classiﬁcation

Semantic Tree model + X

−kernel SVM

Accuracy: 87.6%

(Huang and Kong, 2016)

Pornography classiﬁcation Colour features + CNN Accuracy: 78.29%

(Sigal et al., 2004) Real-time skin segmentation

Bayes’

Detection rate: 86.84%

(Hajraoui and Sabri, 2014)

Face detection Thresholding + Watershed Detection rate 97.27%

(Taqa and Jalab, 2010) Skin detection

ANN Detection rate: 95.62%

(Kim et al., 2017) Skin detection

CNN

Accuracy: 95.62%, Precision: 87.2%, Recall: 91.22%

F-measure: 89.19%

(Zuo et al., 2017) Skin detection

RNN + FCN

AUC (COMPAQ dataset): 95.93%, AUC (ECU dataset): 98.10%

1-EER (COMPAQ dataset): 90.18%, 1-EER (ECU dataset): 94.80%

AUC(%): Area Under Curve

ROC(%): Receiver Operating Characteristic

1-EER(%):Equal Error Rate

MAP: Maximum a Posteriori

SVM: State Vector Machine

COMPAQ: (Bhoyar, 2010)

ometric models to describe the structure of the hu-

man body (Forsyth and Fleck, 1999), or region-based

detection, which extracts local features for recogni-

tion (Hu et al., 2007).

The detection of sexual organs does not, how-

ever, require skin detection as a pre-processing step.

For example, in (Wang et al., 2010), the shape and

colour of the nipple are used to train the Viola-

Jones algorithm with an AdaBoost classiﬁer to de-

tect potential nipple regions. The feature set used by

Wang et al. in the AdaBoost classiﬁer may be ex-

tended to include Haar-like features such as edge, line

and centre-surround features (Lienhart and Maydt,

2002), and colour, texture and shape features obtained

through colour moments, histogram of oriented gra-

dients (HOG) and grey-level co-occurrence matrices

(GLCM) (Kejun et al., 2012).

More elaborate human body models are required

when taking into account other sexual organs. Here,

in addition to the recognition difﬁculties introduced

through pose, the localisation and classiﬁcation prob-

lem should also consider that the genitalia may also

be exposed in coitus (Lv et al., 2011). For exam-

ple, (Shen et al., 2010) use a rudimentary human pose

model based on the skin distribution using the loca-

tion of the face and trunk to aid the classiﬁcation of

the nipple and pube regions.

Table 1 compares the performance of feature-

based sexual organ detectors found in literature.

In light of the remarkable results achieved by

deep learning architectures for various computer vi-

sion tasks, recent advances in pornography detection

are also using deep learning approaches (Moustafa,

2015; Zuo et al., 2017), using convolutional neural

networks (CNN) to learn pornographic features from

image examples (Moustafa, 2015; Zuo et al., 2017).

Such an approach has the advantage of learning the

common global traits in pornographic images. How-

ever, these approaches do not speciﬁcally search for

body parts in the image.

Moustafa describes one of the ﬁrst deep-learning-

based approaches to pornographic image content.

Pre-trained AlexNet and GoogLeNet architectures are

repurposed for pornographic image detection and de-

noted as ANet and GNet respectively. By modify-

ing the third fully-connected layer as a two-way Soft-

max and training on the pornographic dataset (NPDI)

these provide the probabilities that the input image is

pornographic or benign. By considering the two ar-

chitectures as providing complimentary class proba-

bility, Moustafa then adds two new architectures de-

noted as AGNet and AGbNet which fuse the results

from ANet and GNet to improve the ﬁnal classiﬁca-

tion. In AGNet, the fusion score is the average of the

two probabilities while in AGbNet, the fusion score

is the largest of the two. Using a receiver operating

characteristic curve (ROC) curve Moustafa demon-

strates that fusion improves the classiﬁcation perfor-

mance, with AGbNet performing slightly better than

the AGNet (Moustafa, 2015). The use of two convo-

lutional neural networks, however, comes at the cost

of a large number of network parameters that need

training, which in turn, requires a large image dataset

to ensure proper training.

In (Huang and Ren, 2018), the controllability over

the features learnt by the CNN is increased by intro-

ducing a colour feature histogram as an input to the

CNN. This approach achieved an accuracy of 99.31%

on the NPDI data set, outperforming a vanilla CNN by

2.67%. An additional contribution of this work is the

integration of bagging into the CNN to solve the over-

ﬁtting problem and enhance generalisation (Huang

and Ren, 2018). In (Vitorino et al., 2018), the

over-ﬁtting problem is addressed by adopting trans-

fer learning approach, training the GoogLeNet CNN

network ﬁrst on a large dataset of non-pornographic

content and then ﬁne-tuning the network with porno-

graphic content (Vitorino et al., 2018). Table 2 pro-

vides a comparison of deep learning approaches for

pornographic image detection.

Many of the false-positives in pornography clas-

siﬁcation of these deep learning based approaches

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

206

Table 2: Comparison of Deep Learning approaches to pornographic image detection.

Method Aim Approach Performance

(Zuo et al., 2017) Skin detection RNN 98.1% AUC

94.8% 1-EER

(Moustafa, 2015) Pornographic vs AlexNet & GoogLeNet 94.2 ± 2% accuracy

Non pornographic classiﬁcation

(Huang and Ren, 2018) Pornographic vs colour feature histogram & 97.51% accuracy on NPDI dataset

Non pornographic classiﬁcation GoogLeNet CNN

(Vitorino et al., 2018) Detect sexually exploitative imagery of colour feature histogram & 91.5 % accuracy on Pornograpy-2K dataset

children & adults from innocuous images GoogLeNet CNN 86.5% accuracy on a

SEIC dataset

(Wang et al., 2016) Detect exposed body parts CNN with MIL 98.4% accuracy on NPDI dataset

(Jin et al., 2018) region-based recognition GoogLeNet 97.5% accuracy on NPDI

dataset

of sexual organs

(Macedo et al., 2018b) age estimation & ResNet-50 architecture & 79.84% accuracy on

pornography detection VGG-16 for age estimation NSFW

& RCPD datasets

(Perez et al., 2017) video pornography detection CNN GoogLeNet architecture 96.3% accuracy on Porn-2k dataset

fusing MPEG and still image features

MPEG: Moving Picture Experts Group

RCPD: Region-bsed annotated child pornography dataset (Macedo et al., 2018a)

stems from the lack of training on private body

parts (Wehrmann et al., 2018). This issue can be re-

solved by either designing separate classiﬁers for dif-

ferent body parts (Wehrmann et al., 2018) or by us-

ing a generic pornographic content detector trained

to detect the private body parts (Wang et al., 2016;

Jin et al., 2018). In (Wang et al., 2016), a CNN is

trained using multiple instance learning (MIL), thus,

using a generic detector which nevertheless recog-

nises an image as pornographic if it contains at least

one exposed private body part. The results are com-

pared with traditional methods for pornography de-

tection, based on image retrieval and bag-of-features

techniques, as well as with variants of the proposed

method, by either training the CNN on entire images

rather than multiple instances, or training the CNN on

images of the breast and genitalia separately. Wang et

al. report that the generic detector trained on multi-

ple image instances of breasts and genitalia achieved

better pornographic/benign image classiﬁcation.

From this review, we note that although the ap-

proaches described in (Wang et al., 2016) and (Jin

et al., 2018) use image instances with private body

parts to recognise images as pornographic/benign,

there is a lack of deep learning approaches which lo-

cate and assign a label to the different body parts.

However, methods that label instances of body parts

within the image would be of beneﬁt to help law en-

forcement agencies obtain a better understanding of

the content of the image and the reason why the image

is considered pornographic. We address this research

gap in our proposed approach.

3 METHODOLOGY

The aim of this work is the identiﬁcation and labelling

of private body parts, namely, female breasts, and fe-

male and male sexual organs. To achieve this goal,

we propose a two-step process in which the ﬁrst step

classiﬁes the image as pornographic/benign while the

second step performs a more in-depth analysis on

only those images identiﬁed as pornographic, locat-

ing and labelling instances of private body parts. Such

a two-step process allows us to employ the multi-

class classiﬁer to only those images that require fur-

ther analysis. This approach is necessary considering

the speed requirement for law enforcement agencies.

Our pornographic image classiﬁer consists of a CNN,

which takes as its input the entire image. To detect

and label private body parts within the pornographic

image we adopt a windowing approach, whereby each

image is divided into smaller windows, checking each

of these windows with a multi-class CNN.

3.1 Pornography Classiﬁcation

The ﬁrst step of our approach entails the classiﬁcation

of an image as pornographic/benign, thus, a binary

classiﬁer is used. Keeping in mind the speed and ef-

ﬁciency requirements, the lightweight MobileNet ar-

chitecture was adopted as it uses depth-wise separable

convolutions to a build fast, light-weight deep neu-

ral network (Howard et al., 2017). Our evaluations

demonstrate that the accuracy reached by this model

is comparable to that achieved by other models while

having a faster computation time.

We adopt a transfer learning approach on a pre-

trained MobileNet by replacing its classiﬁcation layer

by a custom classiﬁer consisting of a Dense ReLU

Private Body Part Detection using Deep Learning

207

layer followed by a Softmax output layer. Fine tuning

was used to ﬁnd the right balance between using the

pretrained feature extractors and the learning of new

features.

To train this model, we use a subset of the large-

scale pornographic image corpus created by the Uni-

versity of Leon

(UL). This database was generated

by crawling around two million images from ﬁve pop-

ular pornography websites. Since the data set has

1656 benign and 16033 pornographic images, train-

ing using this data introduces a bias. Thus, additional

benign images were introduced from the VOC2012

dataset.

3.2 Private Body Part Classiﬁcation

Once an image is considered to be of a pornographic

nature, it is tessellated into non-overlapping 128 ×

128 windows, passing each of these windows to our

classiﬁer to identify the location, if any, of four differ-

ent body parts, namely buttocks (Bt), female breasts

(FBr), female genitalia (FG), male genitalia (MG).

We note that the appearance of female genitalia can

differ considerably between instances of posing and

instances of sexual activity. Thus, we redeﬁne these

body parts as female genitalia posing (FGP) and fe-

male genitalia active (FGA). This helps the model

learn each class better by reducing the variation in

observations of each object. Moreover, we note that

some pornographic images contain sex toys. While

these may appear similar to genitals, they should not

be labelled as such. Thus, we introduce a sixth class

which corresponds to sex toys (ST). A seventh, benign

(Bn) class is required to describe image content which

is not a private body part. Thus, body-part classiﬁca-

tion requires a multi-class classiﬁer.

To determine the architecture that best suits our

application, multiple pretrained models were trained

for this classiﬁer, training all classiﬁers with image

windows of 128 × 128 pixels. We note that during the

ﬁne tuning stage, the best performance was achieved

when none of layers were excluded from training.

This may be due to the fact that we are working with

windows rather than complete images.

To train the multi-class classiﬁer, points of inter-

est on the pornographic images from the UL dataset

were manually labelled. Five 128 × 128 windows

were taken around each labelled point, one centred

on the labelled point and four perpendicularly offset.

The benign class training set was generated separately

by extracting 128 × 128 windows from random points

http://gvis.unileon.es/dataset/apd-2m/

https://bit.ly/33Jd2rp

Table 3: Window dataset classes and their size.

Class Name Number of windows

Benign (Bn) 9895

Buttocks (Bt) 6490

Female breasts (FB) 19258

Female genitalia posing (FGP) 4865

Female genitalia active (FGA) 7460

Male genitalia (MG) 2395

Sex toys (ST) 1635

on benign images. The seven classes and the corre-

sponding number of samples gathered are shown in

Table 3. From this, we note that some classes have

very few examples. Thus, data augmentation in the

form of image scaling with a scaling factor of up to

0.2, image rotations with rotation angles in the range

of [0,

] and horizontal ﬂips were used to boost the

size of this dataset.

4 RESULTS AND EVALUATION

The results were generated on a Windows computer

running a 64-bit operating system, and featuring an

Intel®Core

i7-8750H CPU at 2.20GHz, 16Gb of

DDR4 memory and a GeForce GTX 1070 graphics

card. Python was used as the development language,

and Tensorﬂow and Keras

as open source packages

to build the networks. Similar to (Vitorino et al.,

2018) and others, we evaluate our classiﬁers by us-

ing the sensitivity, speciﬁcity, F1-score and accuracy

metrics.

The evaluation of the ﬁrst-stage pornographic im-

age classiﬁer, we use a test set of 50 pornographic

and 50 benign images. The ﬁrst row of Table 4 shows

the parameters used to train the MobileNet classiﬁer

which takes 5.178s to load and 0.017s to classify an

image. For the test set used, we obtained a sensitivity

of 0.95 and a speciﬁcity of 0.95 which are satisfactory

for the purpose of our application.

The models used to test the performance of the

private body part classiﬁer are shown in Table 4 along

the parameters of the best version obtained for each.

To compare the performance of these models, we

manually selected windows centred around private

body parts from 50 pornographic images. The F1

score of each was calculated. Since windows might

contain more than one sexual organ, both top-1 and

top-3 accuracy were considered along the computa-

tion and loading time. These values are shown in Ta-

ble 5 and were used to select the most suitable classi-

ﬁer for our intended application.

https://keras.io/

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

208

Table 4: Training Parameters for all the Classiﬁers used.

Training Parameters Classiﬁer Used

Model

Architecture

Batch

Size

Optimiser Epochs Validation

Accuracy

Drop Dense

‘ReLU’

Drop Dense

‘ReLU’

Drop Dense

‘ReLU’

Dense

‘SoftMax’

MobileNet 50

RmsProp

lr=4e-7

25 0.9417 / / / / 0.5 128 2

MobileNet 200

RmsProp

lr=2e-5

50 0.7350 / / 0.4 128 0.4 128 7

VGG-16 300

RmsProp

lr=1e-5

30 0.6860 / / 0.5 128 0.5 128 7

ResNet 150

RmsProp

lr=5e-5

40 0.7663 0.5 128 0.7 128 0.7 128 7

Inception V3 300

RmsProp

lr=4e-5

25 0.7529 / 128 0.4 128 0.4 64 7

Table 5: Model Accuracy and Computation time.

Model Accuracy Computation

time (ms)

F1 score

Top-1 Top-3

ResNet 0.8011 0.9569 9.824 0.7258

Inception-v3 0.8010 0.9731 13.08 0.8011

MobileNet 0.7527 0.9731 7.07 0.7258

VGG-16 0.7097 0.9570 13.02 0.7097

Proving a suspect guilty may entail the process-

ing of large amounts of data in a limited time frame,

hence the speed at which each image is checked is

of utmost importance. Thus, the MobileNet model

was chosen since it performs faster compared to other

models and suffers only from a slight tradeoff in accu-

racy. To better visualise this model’s performance, a

confusion matrix is generated and shown in Figure 2.

The manually selected windows set is made up of la-

belled sexual organs and hence has no benign class.

The majority of misclassiﬁcations are caused by the

male genitalia class which may be attributed to the

limited number of examples used during training in

combination with the resulting relatively large ampli-

ﬁcation of its loss when balancing the weights of each

class.

The whole pipeline is then tested by tessellating

the images identiﬁed as pornographic by the porno-

graphic classiﬁer into windows. The confusion ma-

trix in Figure 2, gives the overall performance of the

private-body part classiﬁcation. A seven-class, pure-

chance classiﬁer would have a detection accuracy of

⁄

. The accuracy achieved by our classiﬁer stands

at 0.62 which is four times better than a random se-

lector. Nevertheless, we note that there is a decline

in the overall detection accuracy when using the full

pipeline. This cutback may be attributed to a misclas-

siﬁcation of benign windows with large skin patches

as private body parts and the misclassiﬁcation of win-

dows containing body parts not centred in the win-

dow. Moreover, the imbalance between classes during

Figure 1: Confusion matrix using whole pipeline.

Figure 2: Confusion matrix with manual windowing.

training was attempted to be catered for by weighting

the error of each class, though this does not compare

to the improvements that can be made by using more

examples of body parts which differ in pose and posi-

tion within the window.

Private Body Part Detection using Deep Learning

209

5 CONCLUSIONS

In this paper we review various classiﬁers that use

both hand-crafted features and deep learning tech-

niques for classifying images as pornographic/benign.

However, no known deep learning approaches were

found to label sexual organs inside these images.

Hence, we propose a two-step approach, which uses

a ﬁrst classiﬁer to classify an image as pornographic

or benign, then tessellates pornographic images into

windows. A second classiﬁer checks these windows

for sexual organs and augments the initial classiﬁca-

tion of the pornographic image by describing the ex-

plicit content in that image.

Our results show that the MobileNet architecture

has similar classiﬁcation performance as the other

networks tested but reduces the computation time by

at least 20%. Thus, it was considered suitable for

our system and achieved an accuracy of 95% in the

ﬁrst step of our pipeline. This result implies that the

pornographic image classiﬁer correctly distinguishes

between benign and pornographic images.

The second step of our pipeline reached an accu-

racy of 72.58% for windows centred around the pri-

vate body parts. This demonstrates that deep learn-

ing approaches can be used to detect and label private

body parts in pornographic images. When evaluating

the full pipeline, an accuracy of 62% was achieved.

While this is larger than a pure chance classiﬁer, we

note that this classiﬁer can be further improved by

having more training data and introducing smarter

window selection schemes. We propose that window

selection can be achieved by extracting a feature map

from the ﬁrst pornography classiﬁer and using this

map to localise regions of interest presumably con-

taining private body parts.

ACKNOWLEDGEMENTS

This research has been funded with support from the

European Commission under the 4NSEEK project

with Grant Agreement 821966. This publication re-

ﬂects the views only of the authors, and the Euro-

pean Commission cannot be held responsible for any

use which may be made of the information contained

therein.

REFERENCES

Bhoyar (2010). Skin color detection model using neural

networks and its performance evaluation. Journal of

Computer Science, 6(9):963–968.

Choi, B., Han, S., Chung, B., and Ryou, J. (2011). Human

body parts candidate segmentation using Laws texture

energy measures with skin color. In 13th International

Conference on Advanced Communication Technology

(ICACT2011), pages 556–560.

Forsyth, D. A. and Fleck, M. M. (1999). Automatic detec-

tion of human nudes. International Journal of Com-

puter Vision, 32(1):63–77.

Fuangkhon, P. and Tanprasert, T. (2005). Nipple detection

for obscene pictures. In Proceedings of the 5th WSEAS

International Conference on Signal, Speech and Im-

age Processing, SSIP’05, pages 242–247.

Hajraoui, A. and Sabri, M. (2014). Face detection algo-

rithm based on skin detection, watershed method and

gabor ﬁlters. International Journal of Computer Ap-

plications, 94:33–39.

Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D.,

Wang, W., Weyand, T., Andreetto, M., and Adam,

H. (2017). Mobilenets: Efﬁcient convolutional neu-

ral networks for mobile vision applications.

Hu, W., Wu, O., Chen, Z., Fu, Z., and Maybank, S. (2007).

Recognition of pornographic web pages by classify-

ing texts and images. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 29:1019–1034.

Huang, L. and Ren, X. (2018). Erotic image recogni-

tion method of bagging integrated convolutional neu-

ral network. In Proceedings of the 2nd International

Conference on Computer Science and Application En-

gineering, CSAE ’18, pages 107:1–107:7, New York,

NY, USA. ACM.

Huang, Y. and Kong, A. W. K. (2016). Using a CNN ensem-

ble for detecting pornographic and upskirt images. In

2016 IEEE 8th International Conference on Biomet-

rics Theory, Applications and Systems (BTAS), pages

1–7.

Jin, X., Wang, Y., and Tan, X. (2018). Pornographic image

recognition via weighted multiple instance learning.

IEEE Transactions on Cybernetics, pages 1–9.

Kejun, X., Jian, W., Pengyu, N., and Jie, H. (2012). Auto-

matic nipple detection using cascaded AdaBoost clas-

siﬁer. In 2012 Fifth International Symposium on Com-

putational Intelligence and Design, volume 2, pages

427–432.

Kim, Y., Hwang, I., and Cho, N. I. (2017). Convolutional

neural networks and training strategies for skin detec-

tion. In 2017 IEEE International Conference on Im-

age Processing (ICIP), pages 3919–3923.

Lienhart, R. and Maydt, J. (2002). An extended set of haar-

like features for rapid object detection. In Proceed-

ings. International Conference on Image Processing,

volume 1, pages I–I. IEEE.

Lv, L., Zhao, C., Lv, H., Shang, J., Yang, Y., and Wang,

J. (2011). Pornographic images detection using high-

level semantic features. In 2011 Seventh International

Conference on Natural Computation, volume 2, pages

1015–1018.

Macedo, J., Costa, F., and dos Santos, J. A. (2018a). A

benchmark methodology for child pornography detec-

tion. In 2018 31st SIBGRAPI Conference on Graph-

ics, Patterns and Images (SIBGRAPI). IEEE.

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

210

Macedo, J., de Oliveira Costa, F., and dos Santos, J. A.

(2018b). A benchmark methodology for child pornog-

raphy detection. 2018 31st SIBGRAPI Conference

on Graphics, Patterns and Images (SIBGRAPI), pages

455–462.

Moustafa, M. N. (2015). Applying deep learning to classify

pornographic images and videos. In 7th Paciﬁc-Rim

Symposium on Image and Video Technology (PSIVT).

Naji, S., Jalab, H. A., and Kareem, S. A. (2018). A survey

on skin detection in colored images. Artiﬁcial Intelli-

gence Review, pages 1–47.

Perez, M., Avila, S., Moreira, D., Moraes, D., Testoni,

V., Valle, E., Goldenstein, S., and Rocha, A. (2017).

Video pornography detection through deep learning

techniques and motion information. Neurocomputing,

230:279 – 293.

Shen, X., Wei, W., and Qian, Q. (2010). A pornographic

image ﬁltering model based on erotic part. In 2010

3rd International Congress on Image and Signal Pro-

cessing, volume 5, pages 2473–2477.

Sigal, L., Sclaroff, S., and Athitsos, V. (2004). Skin color-

based video segmentation under time-varying illumi-

nation. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 26(7):862–877.

Taqa, A. Y. and Jalab, H. A. (2010). Increasing the reliabil-

ity of skin detectors. Scientiﬁc Research and Essays,

5(17):2480–2490.

Tian, C., Zhang, X., Wei, W., and Gao, X. (2018). Color

pornographic image detection based on color-saliency

preserved mixture deformable part model. Multimedia

Tools and Applications, 77(6):6629–6645.

Vitorino, P., Avila, S., Perez, M., and Rocha, A. (2018).

Leveraging deep neural networks to ﬁght child

pornography in the age of social media. Journal

of Visual Communication and Image Representation,

50:303 – 313.

Wang, X., Hu, C., and Yao, S. (2009). An adult image

recognizing algorithm based on naked body detection.

In 2009 ISECS International Colloquium on Comput-

ing, Communication, Control, and Management, vol-

ume 4, pages 197–200.

Wang, Y., Jin, X., and Tan, X. (2016). Pornographic im-

age recognition by strongly-supervised deep multiple

instance learning. In 2016 IEEE International Confer-

ence on Image Processing (ICIP), pages 4418–4422.

Wang, Y., Li, J., Wang, H. L., and Hou, Z. (2010). Auto-

matic nipple detection using shape and statistical skin

color information. In MMM.

Wehrmann, J., Simões, G. S., Barros, R. C., and Cavalcante,

V. F. (2018). Adult content detection in videos with

convolutional and recurrent neural networks. Neuro-

computing, 272:432–438.

Zuo, H., Fan, H., Blasch, E., and Ling, H. (2017). Combin-

ing convolutional and recurrent neural networks for

human skin detection. IEEE Signal Processing Let-

ters, 24(3):289–293.

Private Body Part Detection using Deep Learning

211