Automatic Fracture Detection and Characterization in Borehole Images

Using Deep Learning-Based Semantic Segmentation

Andrei Baraian

, Vili Kellokumpu

, R

aty Tomi

and Leena Kallio

VTT Technical Research Centre of Finland, Kaitov

ayl

a 1, Oulu, Finland

Astrock Oy, Ahventie 4, Espoo, Finland

ﬁ

Keywords:

Semantic Segmentation, Borehole Analysis, DeepLab, Deep Neural Networks.

Abstract:

Fracture analysis represents one of the key investigations that needs to be carried in borehole logs. Identifying

fractures, as well as other similar features (like breakouts or foliations) is essential for characterizing the

reservoir where the drilling took place. However, identifying and characterizing the fractures from borehole

images is a very time and resource consuming task, that require extensive knowledge from geological experts.

For this reason, developing semi-automated or automated tools would facilitate and increase the productivity

of fracture analysis, since even for one reservoir, experts need to analyze and interpret hundreds of meters of

borehole images. This paper presents a deep learning based approach for application of automatic fracture

detection and characterization in borehole images, relying on state-of-the-art convolutional neural network for

accurate semantic segmentation of fractures. Target images consists of color borehole images, as opposed to

acoustic or drill-core images, and uses real world data, both for training the deep learning model and testing

the whole system. The system is evaluated by using multiple metrics and the ﬁnal outputs of the system are

the parameters of the sinusoids that deﬁne the predicted fractures.

1 INTRODUCTION

An important part of borehole geophysics is the anal-

ysis and interpretation of logged measurements for

determining the physical properties of wells or test

holes. Speciﬁc events, such as natural fractures, bed-

dings or foliations approximate the underlying struc-

ture of the soil and offer valuable information, de-

pending on the scope of the borehole. For exam-

ple, micro-fractures in shale allow for identifying the

main route for hydrocarbon, which is a clear indica-

tor of oil or gas reservoirs (Carey et al., 2015). Im-

age well logging is one of the most used technique

for borehole analysis in our days, and can be imple-

mented in several ways, obtaining electrical, acous-

tic, or optical images of the borehole. Another pop-

ular method of analysing fractures is extracting the

drill-core and imaging it at ground-level, using spe-

cialized hardware for rolling the core. Although the

obtained images are quite similar in terms of content,

imaging in borehole is more difﬁcult due to limited

illumination conﬁgurations, poor imaging conditions

in the presence of mud or sediments and prone to im-

age artefacts from camera motion. In our work, the

primary focus is on RGB borehole images captured

by cameras, but it is possible to transfer our algorithm

Figure 1: Borehole imaging concept. Fractures in the 3D

world are represented as planes cutting through a section

of the borehole, generating sinusoid-shaped curves in the

unrolled 2D images.

to acoustic or electrical images (Wedge et al., 2015),

since the images have a high degree of similarity, even

though acquisition methods are different.

To measure and analyse the borehole, a probe

is lowered into the borehole and images are contin-

uously captured with a wide FOV (Field-of-View)

camera. Images can then be unrolled and stitched to-

gether to retrieve a 2D image representation of the in-

ner surface of a borehole. Due to the cylindrical shape

of the borehole, planar features such as fractures and

beddings will be represented as sinusoidal curves in

the unrolled 2D image, as depicted in Fig. 1. Since

one borehole can have several hundreds of meters in

856

Baraian, A., Kellokumpu, V., Tomi, R. and Kallio, L.

Automatic Fracture Detection and Characterization in Borehole Images Using Deep Learning-Based Semantic Segmentation.

DOI: 10.5220/0011673100003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 4: VISAPP, pages

856-863

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

depth, manually analyzing all the images easily be-

comes a laborious and time-consuming task for geol-

ogists. Our proposed solution is an automated system,

capable of accurately and robustly detecting fractures

from borehole images and characterizing them at the

same time in borehole space parameters. Our solution

is developed with data from an existing real world ap-

plication where fractures are manually annotated with

sinusoid parameters. The system is built in such a

way to allow experts to set speciﬁc conﬁdence score

thresholds if more detections are needed.

There are two main approaches for solving frac-

ture detection: traditional image processing tech-

niques combined with pattern recognition, and newer

approaches based on Deep Neural Networks (DNNs)

(Chen et al., 2018). A generic pipeline for fracture

detection consists of two stages: segmenting fracture

pixels from the background pixels, and then approx-

imating the models of the curves describing the frac-

tures. We have identiﬁed the bottleneck of this system

as being the segmentation stage. High quality frac-

ture masks make the ﬁtting easy, however in practice

it is exceptionally hard to perfectly segment fractures

due to their variability in width, texture, contrast and

shape, which forces us to design more complex ﬁt-

ting algorithms. Using traditional image processing

techniques, it is very hard and in some cases impos-

sible to derive the perfect combination of ﬁltering op-

erations capable of robustly segmenting the fractures.

Even methods relying on more advanced algorithms,

such as K-means or kNN clustering are having prob-

lems with deep boreholes where color intensity distri-

butions vary a lot. Nonetheless, recent advances in

deep learning architectures, especially the ones de-

veloped for the task of semantic segmentation, are

proving to achieve remarkable accuracy, robustness

and performance in scene understanding and classiﬁ-

cation, being successfully applied in domains such as

autonomous driving, optical quality inspection, robot

navigation, etc. (Li et al., 2021; Tabernik et al., 2019).

This serves as inspiration in our work for choosing

DeepLabv3+ (Chen et al., 2018) for the semantic seg-

mentation of fractures.

2 RELATED WORK

Detailed characterisation of borehole features, such as

fractures, beddings, foliations, etc. can greatly bene-

ﬁt from the automatization of borehole analysis. Al-

though the research in this domain is still in its begin-

nings, there are several works that have further pushed

the state-of-the-art in borehole analysis and interpre-

tation.

Although the focus in this paper is on optical im-

ages captured with color cameras, techniques devel-

oped for acoustic or resistivity images can be ap-

plied as well. One of the earlier methods proposed

by Thapa et al. (Thapa et al., 1997) was relying on

computing an edge map and searching for sinusoid

candidates by applying the Hough transform. Along

the same direction with subtle differences, (Glossop

et al., 1999), (Zhang and Xiao, 2009), and (Hall et al.,

1996) have applied Hough transform for sinusoid ﬁt-

ting. One of the major drawbacks of HT-based imple-

mentations is the high computational cost since it is a

brute-force method. In (Moran et al., 2020), the au-

thors aim to overcome this limitation by performing

a heuristic search instead of an exhaustive one and

leverage the Iterated Local Search (ILS) algorithm.

Other geometric transformations have also been ap-

plied, such as (van Ginkel et al., 2003), where a gen-

eral Radon transform is applied. These methods are

identifying features that represent fractures, and then

by using the geometric transforms, votes are accu-

mulated from each feature and the parameter values

with highest support are selected to be the sinusoids.

An interesting approach by D. Wedge et al.(Wedge

et al., 2015) ﬁrst highlights regions in the hole most

suitable for automated fracture analysis based on an

image complexity (quality) measure, arguing that im-

ages with a high degree of noise may be ignored even

in the manual labeling process. Their method is im-

plemented both for acoustic televiewer and optical

televiewer, with the only difference of changing the

complexity function.

Recently, methods which leverage the latest ad-

vancements in machine learning (ML) are being suc-

cessfully applied for fracture detection. In (Anatoli

Quintanilla Cruz et al., 2017), although the Hough

transform is still applied for getting all possible candi-

dates, there is an additional step which consists of the

candidates going through a convolutional neural net-

work for validation or exclusion. Dias et al. (Dias

et al., 2020) pave the way towards end-to-end sys-

tems for fracture and breakouts analysis, by utilizing

Fast-RCNN (Girshick, 2015) network to identify frac-

tures and breakouts. To overcome the challenge of

limited data, they generate synthetic data of fractures

and breakouts and train the model using this data. Al-

though they reach a 98% area under ROC curve in

the case of testing on the simulated data, when tested

on real images, the system’s performance drops sig-

niﬁcantly. Moreover, only identifying fractures is not

enough for a complete autonomous system for bore-

hole analysis, as fractures need to be characterized

in terms of borehole parameters such as depth, dip

and dip direction. Developed simultaneously with our

Automatic Fracture Detection and Characterization in Borehole Images Using Deep Learning-Based Semantic Segmentation

857

work, some ideas from (Alzubaidi et al., 2022) are

quite similar to ours, although there are key differ-

ences both in the system design and implementation.

In terms of imaging, they are using unwrapped core

images whereas we use borehole images. To com-

bine the fracture segmentation and instance detection,

they are using Mask R-CNN (He et al., 2017) network

for instance segmentation of fractures. Although the

approach works for the general case and manages to

jointly segment and separate each fracture instance,

some extreme cases where fractures have a high am-

plitude might cause problems. After obtaining the

mask of the fracture, the next step consists in skele-

tonizing the fracture, so that the fracture has a width

of just one pixel, allowing for better and faster curve

ﬁtting. We employ the same idea, however, we per-

form joint separation where multiple fractures might

intersect, enabling us to ﬁt multiple conﬁgurations of

sinusoids, accounting for sinusoids with high ampli-

tudes as well.

3 DATA

The development and testing of this work was car-

ried on three separate datasets, each depicting images

from three different boreholes. The data and anno-

tations are both from a real world application. The

resolution of the images was 720x2000 pixels and 1

meter of the borehole was comprised in one image.

Each borehole had around 200-300 meters in depth,

resulting in 200-300 images per dataset. The anno-

tation of fractures was provided in a separate ﬁle for

each dataset, having the fractures speciﬁed as sinu-

soids in borehole parameters, hence a conversion to

image space was needed in order to generate ground-

truth masks. For each dataset, the data was split in

training, validation and testing splits with a ratio of

60-20-20 % out of all the images present in a dataset.

Consequently, to characterize the predicted sinusoids,

it is necessary to perform the inverse operation, from

image space to borehole parameters. Because frac-

tures are of different widths, empirically a width of 10

pixels was chosen as the default width of the ground-

truth. In practice, the observed fractures are not ex-

actly sinusoids and in borehole images the sinusoid

is often only partially visible. Accurate pixel level

annotations would be ideal for training the semantic

segmentation but annotating data in pixel level is not

feasible in the target application. Therefore we fo-

cus our study to the use of parametric sinusoid labels

produced in the current manual workﬂow of the ap-

plication. We think this is a decent tradeoff, as manu-

ally labelling and correcting fractures would be infea-

sible in real world use. The problem is formulated as

a two-class semantic segmentation problem, with the

classes of fracture and background. Fig. 2 shows im-

ages from the three boreholes from where one can see

the big variation in terms of texture. Other challenges

in the image data include low level artifacts due to

borehole image computation from camera images and

wide vertical line segments where water blocks the

borehole texture as seen in Fig 2.

Figure 2: Example images from boreholes.

4 METHOD

Our proposed method relies on two important stages:

fracture detection using convolutional neural net-

works for semantic segmentation and sinusoid ﬁtting

on detected fractures. The developed system is aimed

at not only detection of fractures, but also accurate

identiﬁcation and characterization of each fracture.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

858

Figure 3: Workﬂow overview. Semantic segmentation and

sinusoid ﬁtting are the two main components of the system.

4.1 Semantic Segmentation

DeepLab is a state-of-the-art Deep Convolutional

Neural Network (DCNN) framework developed by

Chen et al. (Chen et al., 2018) for the task of se-

mantic segmentation in images. The ﬁrst version of

the framework (Chen et al., 2015) proposed the us-

age of atrous convolutions and fully connected Condi-

tional Random Fields (CRF) for overcoming the low

accuracy in accurate object segmentation. By com-

bining the ﬁnal DCNN layer with the fully connected

CRF, the model is able to bypass the invariance prop-

erties of DCNNs, allowing for increased accuracy in

segmentation. Last iteration of the framework (Chen

et al., 2018) introduces a decoder module for the pur-

pose of reﬁning the segmentation results, especially

along the object boundaries. The work is inspired

by combining the advantages from the spatial pyra-

mid pooling module and the encoder-decoder struc-

ture, both heavily used in the semantic segmentation

task. Moreover, the authors adapt the Aligned Xcep-

tion model (Chollet, 2017) for the semantic segmen-

tation task and apply depthwise separable convolution

to both the ASPP and decoder module.

For the training procedure, we have used the pre-

trained Xception model from the authors’ repository

and ﬁne-tuned on our training data, for each dataset.

The Xception backbone is initially pre-trained on

ImageNet-1k dataset (Russakovsky et al., 2015) and

then the DeepLabv3+ model is further pre-trained on

MS-COCO dataset (Lin et al., 2014) and on PAS-

CAL VOC 2012 (Everingham et al., ), for which

the evaluation is also performed. Because we are

more interested in achieving higher accuracy rather

than real-time performance, we drop the experimen-

tation on MobileNetV2 (Sandler et al., 2018) and Mo-

bileNetV3 (Howard et al., 2019). We run the model

with an output stride of 16, hence the atrous rates are

[6, 12, 18]. Due to the large image size, we are train-

ing with a batch size of 1 image. For ﬁne-tuning,

all trained weights are used, except the logits, since

the number of classes is different than in PASCAL

dataset.

Because the density of fractures in a borehole is

quite low, as well as the total number of fractures,

there is a high class imbalance between the fracture

class and the background class. To alleviate this prob-

lem, class weights have been calculated using 1 and

applied in the training procedure.

= N

/(N

cls

∗ N

s j

), (1)

where N

is the total number of samples, N

cls

is the

number of classes and N

s j

represents the number of

samples from class j. The calculated weights are then

used in the model’s loss function.

4.2 Sinusoid Fitting

After getting the masks of possible fractures, the next

goal is to ﬁt sinusoid curves to these masks, so that

fractures are characterised by curve parameters. The

main steps for fracture identiﬁcation consist in the fol-

lowing steps: skeletonization of the predicted frac-

ture masks, separation of fracture intersection points

(joints), connected components labelling of indepen-

dent fracture segments and brute-force sinusoid ﬁt-

ting of multiple independent skeletons. The output

of the semantic segmentation inference consists of a

masked image of the identiﬁed pixels representing a

fracture. However, due to imperfections in segmen-

tation, direct ﬁtting on the predicted masks performs

poorly, both in accuracy and computation time. By

skeletonizing the fracture masks, their width is de-

Automatic Fracture Detection and Characterization in Borehole Images Using Deep Learning-Based Semantic Segmentation

859

(a) (b) (c)

(d) (e)

Figure 4: Processing pipeline. (a) Original image. (b) Semantic segmentation of fractures. (c) Fracture mask is skeletonized,

joints (intersections of fractures) are removed and independent fracture segments are labelled. (d) Fitted sinusoids on the

mask. (e) Sinusoids overlaied on the original image. Blue - GT sinusoids, Green - Detected sinusoids.

creased, however the shape information is still pre-

served and the number of sample points for the sinu-

soid ﬁtting is signiﬁcantly reduced. The skeleton will

have a width of 1 pixel. Furthermore, joints in the

skeleton are identiﬁed and removed, using a simple

heuristic based on the number of neighbours. Joint

points are detected if there are at least 4 neigbours in

an 8-neighborhood region. After the joint points are

removed from the skeletons, a connected component

labeling algorithm is performed, and each individual

skeleton is segmented. Smallest skeleton-fractures

are removed and then a brute-force approach is used

to create pairs of skeleton-fractures. The ﬁtting is per-

formed on the fracture pair, as well as considering

independently each one of the fractures, to account

for the case where only one skeleton-fracture deﬁnes

the whole fracture. To limit the redundant cases, we

match skeleton-fractures with fractures which are not

on the same horizontal level. The ﬁtting problem is

modeled as a simple curve ﬁtting problem in the im-

age space. We assume the case of a vertical borehole,

hence we can approximate the shape of the fractures

with a sinusoid curve, having the following equation:

y = A ∗ sin(B ∗ (x + H))+V, (2)

where A is the amplitude of the sinusoid,

2∗π

is the pe-

riod, although for our case we know that the fractures

are sinusoids having only one period, so B=1 for our

case, H is the phase shift and V is the (depth) verti-

cal shift. The ﬁtting is performed by using SciPy’s

non-linear least squares optimization library (Virta-

nen et al., 2020). We are using the Levenberg-

Marquardt algorithm without any bounds added. As

in any unconstrained optimization problem, we need

to estimate an initial guess. For the amplitude and

phase shift, we initialise them with a value of 1 and

for the vertical shift, we calculate the histogram of the

mask on columns and pick the location of the maxi-

mum value as the ﬁrst guess for the vertical shift.

Once the ﬁtting is done and all plausible sinu-

soids have been computed, we run a Non-Maximal

Suppression (NMS) on the detected sinusoids to dis-

card overlapping instances. The detected sinusoid has

only 1 pixel width, hence it’s width is increased by a

ﬁxed amount to account for robustness and overlayed

on the initial masks generated by the segmentation,

from where a conﬁdence score is derived, based on

the overlapping between the detected sinusoid and the

generated mask. Then, in a greedy way, only the best

sinusoids are retained.

Ultimately, the detected sinusoids parameters are

converted from the image-space into the borehole

space, in terms of dip angle, dip direction and frac-

ture depth. To calculate the dip angle, the following

equation is used:

ω = arctan(

− D

), (3)

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

860

where ω is the dip angle, D represents the diameter

of the borehole, D

is the depth of the upper angle

and D

is the depth of the lower angle. The dip di-

rection is easily computed by identifying the smaller

angle of the amplitudes and converting it from pixels

to degrees.

5 EXPERIMENTAL RESULTS

To assess the performance of our system, a multi-

tude of metrics have been applied, to ensure reli-

able measurements. The performance of the semantic

segmentation model can be established by using the

mIoU metric, however this is heavily dependent on

the ground-truth masks, which were obtained by plot-

ting the sinusoids with a ﬁxed width.

Figure 5: Precision and recall for the 3 datasets.

In practice, fractures are not perfectly following a

sinusoidal shape, but there can be slight deviations.

Moreover, the width of real fractures varies and es-

tablishing such an accurate ground-truth is a labori-

ous work, which is beyond the scope of this work.

Hence, we used the mIoU just a mere indication of

how to choose some training parameters, such as label

weights, atrous rates or batch size. To get an insight

on how key parameters are inﬂuencing the results, the

Precision-Recall curve is an excellent instrument that

allows the user to select the optimal tradeoff in terms

of sensitivity and speciﬁcity.

To derive the PR curve, we ﬁrst need to classify all

detections and ground-truth instances into true pos-

itives (TP), false positives (FP). Additionally, each

detection needs a conﬁdence score, so that we can

choose the optimal tradeoff. We deﬁne the conﬁ-

dence score as the number of inliers obtained from

Figure 6: The results of dip and dip direction calculation for

the three datasets.

overlaying the detected sinusoid on the ground-truth

and counting the common pixels. From a computer

vision perspective, it would be enough to classify the

detections based on the IoU score, however, given that

our end goal for the system is to deliver sinusoid pa-

rameters, it is more desirable to compare the detec-

tions with the ground-truth in the borehole reference

system, namely comparing fractures’ depth, dip angle

and dip direction by computing the absolute error. In

Fig. 5, the PR curve is computed for multiple error

thresholds, expressed in absolute units (5-30%). This

has been done to overcome the limitation of generat-

ing a ﬁxed width for all fractures. However, we can

select a decent error margin of 20 units and then cal-

culate the confusion matrix, which can be inspected

in Table 1. After carefully analyzing the results de-

picted in Fig. 6, we came to the same conclusion as

(Alzubaidi et al., 2022), that dip direction results are

less accurate than dip angle or depth of fracture re-

Automatic Fracture Detection and Characterization in Borehole Images Using Deep Learning-Based Semantic Segmentation

861

(a) . (b) .

Figure 7: Sinusoid ﬁtting results. Color scheme: Blue - ground-truth (GT) sinusoid, detected; Green - predicted sinusoid,

matched with GT; Red - GT sinusoid, not detected; Pink - predicted sinusoid, not matched with GT.

sults, due to being highly sensitive to the sinusoid ﬁt-

ting and fracture segmentation. Moreover, in the case

of the confusion matrix, if one sinusoid prediction is

off by more than a certain percentage, then it will pro-

duce both a false positive and a false negative. Un-

fortunately, due to the restricted application domain,

there are no open-datasets available to benchmark our

method.

Table 1: Confusion matrix for a threshold of 20 units.

Datasets TP FP FN

Dataset#1 66 39 58

Dataset#2 55 17 36

Dataset#3 89 50 89

6 CONCLUSION

This work paves the way towards a fully automatic

system for fracture recognition and characterization

in borehole images. The proposed workﬂow trains

a semantic segmentation network by using only real

images and performs inference to detect possible

fracture pixels. The second stage of the workﬂow ﬁts

sinusoid curves on the detected fractures and retains

the best matches, ultimately delivering as output the

sinusoid parameters in the borehole space parameters,

namely fracture depth, dip angle and dip direction.

Additionally, we output a conﬁdence score that can be

used in further ﬁltering of the detections by a human

operator. The developed algorithm is accurate and

faster than a human interpretation, hence representing

a major help for geologists. As a future work, we aim

to benchmark other DNN for semantic segmen-

tation and improve the robustness of the algorithm.

ACKNOWLEDGEMENTS

This work was supported by Real-Time AI-Supported

Ore Grade Evaluation for Automated Mining - RAGE

project. We also gratefully acknowledge the sup-

port of ASTROCK Oy for providing the borehole im-

ages and the corresponding ground-truth annotations.

The work is part of the Academy of Finland Flag-

ship Programme, Photonics Research and Innovation

(PREIN), decision 320168.

REFERENCES

Alzubaidi, F., Makuluni, P., Clark, S. R., Lie, J. E.,

Mostaghimi, P., and Armstrong, R. T. (2022). Auto-

matic fracture detection and characterization from un-

wrapped drill-core images using mask r–cnn. Journal

of Petroleum Science and Engineering, 208:109471.

Anatoli Quintanilla Cruz, R., Cacau, D. C., dos Santos,

R. M., Ribeiro Pereira, E. J., Leta, F. R., and Gonza-

lez Clua, E. (2017). Improving accuracy of automatic

fracture detection in borehole images with deep learn-

ing and gpus. In 2017 30th SIBGRAPI Conference

on Graphics, Patterns and Images (SIBGRAPI), pages

345–350.

Carey, J. W., Lei, Z., Rougier, E., Mori, H., and

Viswanathan, H. (2015). Fracture-permeability be-

havior of shale. Journal of Unconventional Oil and

Gas Resources, 11:27–43.

Chen, L., Papandreou, G., Kokkinos, I., Murphy, K., and

Yuille, A. L. (2015). Semantic image segmentation

with deep convolutional nets and fully connected crfs.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

862

In Bengio, Y. and LeCun, Y., editors, 3rd Interna-

tional Conference on Learning Representations, ICLR

2015, San Diego, CA, USA, May 7-9, 2015, Confer-

ence Track Proceedings.

Chen, L., Zhu, Y., Papandreou, G., Schroff, F., and Adam,

H. (2018). Encoder-decoder with atrous separable

convolution for semantic image segmentation. In Fer-

rari, V., Hebert, M., Sminchisescu, C., and Weiss, Y.,

editors, Computer Vision - ECCV 2018 - 15th Euro-

pean Conference, Munich, Germany, September 8-14,

2018, Proceedings, Part VII, volume 11211 of Lecture

Notes in Computer Science, pages 833–851. Springer.

Chollet, F. (2017). Xception: Deep learning with depth-

wise separable convolutions. In 2017 IEEE Confer-

ence on Computer Vision and Pattern Recognition,

CVPR 2017, Honolulu, HI, USA, July 21-26, 2017,

pages 1800–1807. IEEE Computer Society.

Dias, L. O., Bom, C. R., Faria, E. L., Valent

ın, M. B., Cor-

reia, M. D., de Albuquerque, M. P., de Albuquerque,

M. P., and Coelho, J. M. (2020). Automatic detection

of fractures and breakouts patterns in acoustic bore-

hole image logs using fast-region convolutional neu-

ral networks. Journal of Petroleum Science and Engi-

neering, 191:107099.

Everingham, M., Van Gool, L., Williams, C.

K. I., Winn, J., and Zisserman, A. The

PASCAL Visual Object Classes Challenge

2012 (VOC2012) Results. http://www.pascal-

network.org/challenges/VOC/voc2012/workshop/index.html.

Girshick, R. (2015). Fast r-cnn. In International Conference

on Computer Vision (ICCV).

Glossop, K., Lisboa, P. J. G., Russell, P. C., Siddans, A., and

Jones, G. R. (1999). An implementation of the hough

transformation for the identiﬁcation and labelling of

ﬁxed period sinusoidal curves. Comput. Vis. Image

Underst., 74:96–100.

Hall, J., Ponzi, M., Gonfalini, M., and Maletti, G. (1996).

Automatic Extraction And Characterisation Of Geo-

logical Features And Textures Front Borehole Images

And Core Photographs. volume All Days of SPWLA

Annual Logging Symposium. SPWLA-1996-CCC.

He, K., Gkioxari, G., Doll

ar, P., and Girshick, R. (2017).

Mask r-cnn. In 2017 IEEE International Conference

on Computer Vision (ICCV), pages 2980–2988.

Howard, A., Sandler, M., Chu, G., Chen, L.-C., Chen, B.,

Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V.,

Le, Q. V., and Adam, H. (2019). Searching for mo-

bilenetv3. In ICCV.

Li, J., Jiang, F., Yang, J., Kong, B., Gogate, M., Dashtipour,

K., and Hussain, A. (2021). Lane-deeplab: Lane se-

mantic segmentation in automatic driving scenarios

for high-deﬁnition maps. Neurocomputing, 465:15–

25.

Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ra-

manan, D., Doll

ar, P., and Zitnick, C. L. (2014). Mi-

crosoft coco: Common objects in context. In Fleet,

D., Pajdla, T., Schiele, B., and Tuytelaars, T., edi-

tors, Computer Vision – ECCV 2014, pages 740–755,

Cham. Springer International Publishing.

Moran, M. B. H., Cuno, J. S., Riveaux, J. A., Vasconcellos,

E. C., Biondi, M., Clua, E. W., Correia, M. D., and

Conci, A. (2020). Automatic sinusoidal curves detec-

tion in borehole images using the iterated local search

algorithm. In 2020 International Conference on Sys-

tems, Signals and Image Processing (IWSSIP), pages

255–260.

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh,

S., Ma, S., Huang, Z., Karpathy, A., Khosla, A.,

Bernstein, M., Berg, A. C., and Fei-Fei, L. (2015).

ImageNet Large Scale Visual Recognition Challenge.

International Journal of Computer Vision (IJCV),

115(3):211–252.

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and

Chen, L.-C. (2018). Mobilenetv2: Inverted residuals

and linear bottlenecks. In CVPR.

Tabernik, D.,

Sela, S., Skvar

c, J., and Sko

caj, D. (2019).

Segmentation-Based Deep-Learning Approach for

Surface-Defect Detection. Journal of Intelligent Man-

ufacturing.

Thapa, B., Hughett, P., and Karasaki, K. (1997). Semi-

automatic analysis of rock fracture orientations from

borehole wall images. Geophysics, 62(1):129–137.

cited By 32.

van Ginkel, M., Kraaijveld, M. A., van Vliet, L. J., Reding,

E. P., Verbeek, P. W., and Lammers, H. J. (2003). Ro-

bust curve detection using a radon transform in orien-

tation space. In Big

un, J. and Gustavsson, T., editors,

Image Analysis, 13th Scandinavian Conference, SCIA

2003, Halmstad, Sweden, June 29 - July 2, 2003, Pro-

ceedings, volume 2749 of Lecture Notes in Computer

Science, pages 125–132. Springer.

Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M.,

Reddy, T., Cournapeau, D., Burovski, E., Peterson, P.,

Weckesser, W., Bright, J., van der Walt, S. J., Brett,

M., Wilson, J., Millman, K. J., Mayorov, N., Nel-

son, A. R. J., Jones, E., Kern, R., Larson, E., Carey,

C. J., Polat,

I., Feng, Y., Moore, E. W., VanderPlas,

J., Laxalde, D., Perktold, J., Cimrman, R., Henriksen,

I., Quintero, E. A., Harris, C. R., Archibald, A. M.,

Ribeiro, A. H., Pedregosa, F., van Mulbregt, P., and

SciPy 1.0 Contributors (2020). SciPy 1.0: Fundamen-

tal Algorithms for Scientiﬁc Computing in Python.

Nature Methods, 17:261–272.

Wedge, D., Holden, E.-J., Dentith, M., and Spadaccini,

N. (2015). Automated structure detection and anal-

ysis in televiewer images. ASEG Extended Abstracts,

2015(1):1–4.

Zhang, X. and Xiao, X. (2009). Detection of fractures in

borehole image. In Zhang, T., Hirsch, B., Cao, Z.,

and Lu, H., editors, MIPPR 2009: Automatic Target

Recognition and Image Analysis, volume 7495, pages

1043 – 1048. International Society for Optics and Pho-

tonics, SPIE.

Automatic Fracture Detection and Characterization in Borehole Images Using Deep Learning-Based Semantic Segmentation

863