Pattern Recognition 51 (2016) 242–254
Contents lists available at ScienceDirect
Pattern Recognition
journal homepage: www.elsevier.com/locate/pr
Detection of bird nests in overhead catenary system images
for high-speed rail
Xiao Wu a,n, Ping Yuan a, Qiang Peng a, Chong-Wah Ngo b, Jun-Yan He a
a School of Information Science and Technology, Southwest Jiaotong University, No. 111, North Section 1, 2nd Ring Road, Chengdu, China
b Department of Computer Science, City University of Hong Kong, #83, Tat Chee Avenue, Kowloon, Hong Kong
a r t i c l e i n f o
a b s t r a c t
Article history:
Received 5 September 2014
Received in revised form
13 June 2015
Accepted 13 September 2015
Available online 28 September 2015
Keywords:
Bird nest detection
Image classification
Overhead catenary system
High-speed rail
Intelligent transportation system
The high-speed rail system provides a fast, reliable and comfortable means to transport large number of
travelers over long distances. The existence of bird nests in overhead catenary system (OCS) can hazard
to the safety of the high-speed rails, which will potentially result in long time delays and expensive
damages. A vision-based intelligent inspection system capable of automatic detection of bird nests built
on overhead catenary would avoid the damages and increase the reliability and punctuality, and
therefore is attractive for a high-speed railway system. However, OCS images exhibit great variations
with lighting changes, illumination conditions and complex backgrounds, which pose great difficulty for
automatic recognition. This paper addresses the problem of automatic recognition of bird nests for OCS
images. Based on the unique properties of bird nests, we propose a novel framework, which is composed
of five steps: adaptive binarization, trunk/branch detection, hovering point detection, streak extraction
and pattern learning, for bird nest detection. Two histograms, Histogram of Orientation of Streaks (HOS)
and Histogram of Length of Streaks (HLS), are novelly proposed to capture the distributions of orienta-
tions and lengths of detected twig streaks, respectively. They are modeled with Support Vector Machine
to learn the patterns of bird nests. Experiments on different high-speed train lines demonstrate the
effectiveness and efficiency of the proposed work.
& 2015 Elsevier Ltd. All rights reserved.
1.
Introduction
High-speed trains are well developed around the world due to
their numerous advantages. Indeed, they are safe, reliable, sus-
tainable, convenient and comfortable for passengers. Over the past
few years, China's high-speed rail (HSR) network has progressively
expanded and now become the world's longest high-speed rail
network with around 9300 km of routes [1]. With the fast devel-
opment of high-speed rail network, railway officials and managers
are facing arduous tasks to ensure that the high-speed rail system
is operating in an orderly and reliable way. Among them, safety
has the highest priority for high-speed rail system, especially after
a fatal high-speed railway accident happened near Wenzhou,
China on July 23, 2011, which has caused great concerns on the
safety of high-speed rail network.
The Pantograph–Catenary (PAC) system is the dominant form for
supplying the vital power to railway electrical trains. A pantograph is
an apparatus mounted on the roof of an electric train to collect power
through contact with an overhead catenary equipment called the
n Corresponding author.
E-mail address: wuxiaohk@home.swjtu.edu.cn (X. Wu).
http://dx.doi.org/10.1016/j.patcog.2015.09.010
0031-3203/& 2015 Elsevier Ltd. All rights reserved.
Overhead Catenary System (OCS) [2]. The steel rails on the tracks act
as the electrical return. The OCS is a high voltage system consisting of
contact wire and catenary wire suspended via supports primarily on
poles placed along the railway. The OCS includes messenger wire,
contact wire, droppers, and supporting structure, which consists of
metallic poles, cross-arms, and running rails. The structure of OCS is
illustrated in Fig. 1. In order to achieve good current collection, the
contact wire has to be placed geometrically within defined limits,
which is usually achieved by supporting the contact wire from above
by a second wire known as the messenger wire or catenary. This wire
is attached to the contact wire at regular intervals by vertical wires
known as droppers or drop wires. The messenger wire is supported
regularly at structures, by a pulley, link, or clamp. Due to space lim-
itation, we will not elaborate the details of each component.
Despite offering a balance in cost effectiveness and system
reliability, the important railway power supply chain is a major
cause of train failure faults. The defects in OCS, for example, down
hanging, ripped off droppers, bondings, broken insulators and bird
nests, can result in long time delays, expensive damages, and even
disasters. The existence of bird nests greatly threats the safety of
the high-speed rail network. It will lead to a defective uptake of
energy by the locomotive, resulting in significant energy loss and
damage to the overhead contact wire and pantograph. While this
X. Wu et al. / Pattern Recognition 51 (2016) 242–254
243
and overlapped. Moreover, discerning catenary system from
complex background such as mountains, trees and buildings, is
generally difficult. Therefore, it is extremely challenging for an
automatic OCS image inspection system to function as expected
under various practical concerns. Several representative OCS
images with bird nests are demonstrated in Fig. 2, while some
difficult examples are also shown in Fig. 10.
In this paper, we explore an automatic detection of bird nests
for high-speed rails, which analyzes the captured OCS images and
assists technicians to make decisions. Experiments demonstrate
that the proposed approach achieves a promising performance for
bird nest recognition. The contributions of this paper are as
follows:
To the best of our knowledge, this is the first work to system-
atically analyze the properties of bird nests in OCS images, and
the first to automatically detect bird nests by image processing
technology for OCS inspection in high-speed rail system.
Based on the features of bird nests in OCS system, a five-phase
framework is novelly proposed to detect bird nests in image
sequences, which includes binarization,
trunk and branch
identification, hovering point detection, streak extraction and
pattern learning.
To model the properties of unordered, non-parallel, distributed
and diverse twigs of bird nests, HOS and HLS histograms are
proposed to represent the distributions of orientations and
lengths of detected streaks, which are exploited to detect the
presence of bird nests.
Experiments on multiple sequences of images from real high-
speed rail lines demonstrate the effectiveness and efficiency of
the proposed approach.
This paper is organized as follows. Section 2 gives a brief
overview of related work. Section 3 introduces the system archi-
tecture of the OCS inspection system, and the proposed framework
of automatic bird nest detection system. Section 4 elaborates the
detailed process of bird nest detection. Section 5 describes the
experimental setup and empirical results. Finally, Section 6 con-
cludes this paper.
2. Related work
2.1. Pantograph-Catenary inspection system
to monitor
Conventional OCS inspection systems use physical instruments
mounted on an inspection vehicle to measure the status of pan-
tograph and catenary. The Fiber Bragg Grating (FBG) sensors on a
pantograph are used in [3]
the underground
pantograph-catenary system, which measure the contact force and
the vertical acceleration of the pantograph head. An automatic
diagnostic system is installed in a special trolley running up to
100 km/h to check the health conditions of catenary [4]. Particular
considerations are given on the values and the trends of voltage,
height, stagger, wear, forces and so on. These instruments are
mounted on the pantograph itself and signal wires are attached to
each instrument [4,5], which affect the dynamic characteristics of
the pantograph. An optical radar system [6] is equipped on the
inspection vehicle to record catenary related parameters like
contact wire position, wire wear, pole position as well as distance
between pole and track. However, these physical inspection sys-
tems mainly focus on checking the status and properties of
catenary systems, which are not applicable for bird nest detection.
With the recent development of computerized image recogni-
tion, current works begin to focus on video based detection to
discover defects in OCS [7–15], taking advantage of the video
Fig. 1. Structure of overhead catenary system of high-speed rail.
is a problem common to all trains of electric traction, it is espe-
cially critical for high speed tracks, since the reliability of power
collection decreases as train speed increases. It is the permanent
aim of all railway line operators to detect faults at an early stage,
and to correct any damage, fault or wear on time in order to
prevent serious disturbances of railway traffic. Therefore, it is
greatly desired that intelligent approaches can be designed and
implemented to provide support for technicians. In this paper, we
will focus on the automatic detection of bird nests for high-
speed rails.
The demand for safety and high reliability grows with the
development of high-speed railway lines, leading to the invention
of new monitoring devices. The OCS inspection system is one of
the key items for high-speed rail system. Traditionally,
it is
implemented by laser scanner on top of the train. Compared to the
extremely expensive laser scanner, a breakthrough video mon-
itoring system with on-board high-resolution cameras equipped in
the high-speed trains becomes a promising solution due to its high
functionality, flexibility and interoperability. It can directly capture
the OCS images/videos to ensure the reliability of the OCS. Con-
ventionally, the images/videos have to be visually checked and
evaluated offline by trained technicians, often frame by frame, to
inspect the status of the OCS. It is time-consuming, laborious, and
impractical to manually monitor millions of overhead catenary
supporters along thousand miles of railways. In this work, we
deploy an OCS image inspection system featured with non-contact
cameras for high-speed trains.
Unfortunately, we notice that OCS images exhibit great varia-
tions in lighting changes,
illumination conditions, occlusion,
complex structures, and mixture of foreground and background,
which make automatic detection of bird nests a challenging task.
First, the images from different train lines are captured with varied
weather, lighting changes, viewpoints and illumination conditions.
Second, bird nests exist at different locations of the OCS. And the
regions having bird nests are relatively small and not easy to be
noticed. Third, the overhead catenary structures are complex with
messy crossing lines, especially when the foreground and back-
ground are mixed together, so that many lines are crossed over
244
X. Wu et al. / Pattern Recognition 51 (2016) 242–254
Fig. 2. Examples of bird nests appearing at different locations of the overhead catenary system of high-speed rail. The image qualities are suffered from different variations in
lighting changes, complex structures, mixture of foreground and background, which make automatic detection of bird nests a challenging task.
camera that offers higher resolution and lower cost. The most
related works to our research are [8,9]. With IRIS 320 inspection
vehicle in France [8],
image analysis strategy is adopted for
catenary inspection for preventive maintenance. It locates the
catenary sections between two supporting arms, and extracts all
surrounding elements such as contact wire, carrying wire, sup-
porting arms and simple droppers. The work [9] deals with the
automatic recognition of the catenary elements, such as support-
ing arms, droppers, and droppers with electrical connection. It
consists of two steps: a classical scene analysis is conducted to
identify elements by segmentation in vertical and horizontal
components, followed by feature extraction and classification.
Next, the results are checked and analyzed by analyzing the
catenary element sequence using a Markov model. In [11], a video-
based obstacle detection system is developed in Germany, which
automatically detects obstacles in the pantograph gauge and
retracts the pantograph before collision with an obstacle. With the
recognition of steady arms,
the system is
improved considerably. A system for the automatic detection of
droppers in catenary staves is proposed in [13]. Based on a
the reliability of
X. Wu et al. / Pattern Recognition 51 (2016) 242–254
245
top-down approach, the system exploits priori knowledge to
perform reliable extraction of droppers. In [10], the problematic of
contact wire wear in railways is first presented and a computer
vision system is studied to improve the precision of the systems
for wear measurement of contact wire. A sensor system with a
line-scan camera is used to monitor the interaction between the
catenary and the pantograph [14], which can detect the occurrence
of defects in the catenary-pantograph interaction. In [15], the
pantograph-overhead contact wire system is investigated by using
an infrared camera. In order to detect the temperature along the
strip from a sequence of infrared images, a segment-tracking
algorithm based on the Hough transform is employed. It helps
maintenance operations in the case of overheating of the panto-
graph strip, bursts of arcing, or an irregular positioning of the
contact line. Although the aforementioned works focus on iden-
tification of the components and defects for OCS systems, direct
comparison with them is not possible because these works are not
tailored for bird nest detection. To the best of our knowledge, few
research works and systems are dedicated to the bird nest detec-
tion for the topic of automatic video based diagnostics at
catenaries.
2.2.
Image classification and abnormal pattern detection
Bird nest detection for OCS images belongs to the research
issues of image classification and object detection, which are key
research areas in computer vision and image processing. Image
classification [16] and object detection [17,18] have been exten-
sively studied for decades and been applied to different areas, such
as adverse weather detection [19], human detection [20–22], and
intelligent transportation [23,24]. An image is classified according
to its visual content, e.g., the existence of bird nests or not.
Element-independent features such as color, shape, texture, gra-
dient or contour [16] are extracted from the image in the first
place. Then in the recognition process, these features are gathered
and knowledge is incorporated in order to identify the objects.
Recently, an object detection system is described in [17], which
represents highly variable objects using mixtures of multiscale
deformable part models. These models are trained using a dis-
criminative procedure that only requires bounding boxes for the
objects in a set of images. The system attains state-of-the-art
results in terms of efficiency and accuracy. In order to localize and
segment objects, R-CNN is proposed in [18], which applies high-
capacity convolutional neural networks (CNNs) for bottom-up
region proposals. In addition, supervised pre-training on a large
auxiliary dataset, followed by domain-specific fine-tuning on a
small dataset, is an effective paradigm for learning high-capacity
CNNs when data are scarce. This method yields a significant per-
formance boost compared to state-of-the-art technologies. A
compositional model for human detection is built in [20] by
exploiting the analogy between human body and text. A dis-
criminative alphabet is automatically learnt to represent body
parts. Based on this alphabet, the flexible structure of human body
is expressed by means of symbolic sequences, which correspond to
various human poses and allow for robust and efficient matching.
Experiments on standard benchmarks demonstrate that the pro-
posed algorithm achieves state-of-the-art or competitive perfor-
mance. A simple yet effective detector for pedestrian detection is
proposed in [21], which incorporates common sense and everyday
knowledge into the design of simple and computationally efficient
features. A statistical model of the up-right human body is
deployed where the head, the upper body, and the lower body are
treated as three distinct components, from which a pool of rec-
tangular templates is tailored to this shape model. Since different
kinds of low-level measurements are incorporated, the resulting
multi-modal and multi-channel Haar-like features represent
characteristic differences between parts of the human body, which
are robust against variations in clothing or environmental settings.
To identify bird nests in an image of OCS scene, previously
explored works on image classification and object detection could
be beneficial for our study. To represent the distinct patterns of
various applications, different kinds of histograms are extracted
and trained for image classification. A system based on computer
vision is presented in [19] to detect the presence of rain or snow. A
histogram of orientations of rain or snow streaks is computed with
the method of geometric moments, which is assumed to follow a
model of Gaussian uniform mixture. The orientation of the rain or
the snow is represented with Gaussian distribution whereas the
orientation of the noises is represented with uniform distribution.
Expectation maximization (EM) is used to separate these two
distributions. An image-based vehicle-type recognition is pro-
posed in [24], which uses Gabor wavelet transform and the Pyr-
amid Histogram of Oriented Gradients (PHOG) features. A reliable
classification scheme is proposed by cascade classifier ensembles.
A novel approach for near-duplicate keyframe identification is
proposed in [25] by matching, filtering and learning of local
interest points. Owing to the robustness consideration,
the
matching of local points across keyframes forms vivid patterns.
Pattern entropy is proposed to capture the matching patterns with
the histogram of matching orientation, and then learn the patterns
with SVM for discriminative classification. Although the existing
approaches on image classification may not be directly applicable
for bird nest detection of OCS images, the idea of capturing the
patterns in the form of histograms enlightens this research. In [26],
a robust abnormal event detection framework based on sparse
reconstruction over the normal bases is proposed. Given a col-
lection of normal training examples, the sparse reconstruction cost
(SRC) is proposed to measure the normalness of the testing sam-
ple. By introducing the prior weight of each basis during sparse
reconstruction, the proposed measurement is more robust com-
pared to other outlier detection criteria. An approach is proposed
in [27] to detect aberrations in video streams using Entropy. It is
estimated on the statistical treatments of the spatiotemporal
information of a set of interest points within a region of interest,
by measuring their degree of randomness of both directions and
displacements. A framework is proposed in [28] to robustly
identify local motions of interest in an unsupervised manner by
taking advantage of group sparsity. In order to robustly classify
action types,
local motion is emphasized by combining local
motion descriptors and full motion descriptors, and then group
sparsity is applied to emphasize motion features using the mul-
tiple kernel method. Overall, there is no prior works on bird nest
detection system presently available. The feasibility and effec-
tiveness of the commonly used features and classification solu-
tions are worth exploration.
3. System architecture and framework
In this section, we will first present the system architecture of
the adopted OCS inspection system, followed by an introduction of
the properties of bird nest. Finally, we will describe the proposed
framework for the automatic detection of bird nests.
3.1. Architecture of OCS inspection system
The architecture of OCS inspection system is shown in Fig. 3.
The inspection system is equipped with two on-board CCD cam-
eras with resolutions of 2456 2058 (5 million pixels) and
1392 1040 (1 million pixels), respectively. The high resolution
camera points toward the catenary to capture the images of OCS,
while the low resolution one focuses on the supporter to capture
246
X. Wu et al. / Pattern Recognition 51 (2016) 242–254
Fig. 3. System architecture of OCS inspection system.
OCS Images
Adaptive
Binarization
Trunk and Branch
Detection
Hovering Point
Detection
Twig Streak
Extraction
Pattern Learning
and Detection
Bird Nest Images
Fig. 4. System framework of the intelligent bird nest detection.
insulators, and supporters.
the pole ID and milestones, so that the technicians can locate the
position of the supporters having bird nests. In this work, we only
consider the OCS images captured with the high resolution cam-
era. The images usually include the content of contact wires,
messenger wires,
In order to best
capture the OCS components in the middle of an image, the
cameras are required to point directly toward the OCS structure.
This device is placed in high-speed trains along the direction of the
movement with a certain angle to capture the still images of the
OCS, while the train is running with a maximum speed of 350 km/
h. The two cameras simultaneously take photos with a rate of 17
frames per second. It produces a large number of images in each
examination. The on-board inspection device is connected with a
portable hard disk to store the captured images. The portable hard
disk can be conveyed and connected to the offline PC for data
analysis and processing, including automatic detection of bird
nests, which is the focus of our work.
3.2. Properties of bird nest
We will first study the properties of bird nest. A bird nest is the
place where a bird incubates its eggs and raises its youngs. It
popularly refers to a specific structure made by the bird itself. For
the bird nests in high-speed rail, the grassy cup nest is the most
common type. The cup nest is smoothly hemispherical inside, with
a deep depression to house the eggs. It is made of pliable mate-
rials, mainly grasses and plant fibers. In flight, the bird breaks a
small twig from a tree and presses it into the saliva, angling the
twig downwards, so that the central part of the nest is the lowest.
It continues adding globs of saliva and twigs until it has made a
crescent-shaped cup [29]. By our observation, bird nest demon-
strates three prominent aspects:
Bird nest is usually constructed on a foundation with strong and
solid basis, such as the supporter or the portal structure in
the OCS.
A bird nest is composed of a bunch of short or long twigs with
the
unordered, non-parallel organization.
groups of streaks belonging to a bird nest should be with
distributed and diverse orientations.
At the same time, the groups of streaks of a bird nest should be
In other words,
with inconsistent lengths.
Motivated by the properties of bird nest, we propose a novel
framework for bird nest detection.
3.3. Framework of the proposed approach
Based on these features of a bird nest, the proposed approach
mainly consists of five phases: pre-processing, trunk and branch
identification, hovering point detection, twig streak extraction,
and pattern learning of streaks. The purpose of these phases is to
verify the presence of bird nest. The system framework of the
proposed approach is illustrated in Fig. 4. To smooth the back-
ground and highlight the discrimination between background and
catenary areas, an adaptive binarization step is first conducted.
Trunk and streak identification is then performed to separate the
trunks that the bird nest will be hosted and the detailed branches
X. Wu et al. / Pattern Recognition 51 (2016) 242–254
247
that are the potential twigs of a bird nest. The hovering points are
detected so that the twigs can be located. After that, the streaks
are extracted using Hough transform to convert the twigs to
straight lines. Two histograms, so-called HOS and HLS, are built by
accumulating the orientations and lengths of different compo-
nents obtained by the hovering point detection. The distributions
of HOS and HLS are then modeled using SVM to discriminate the
potential regions with or without a bird nest. A decision criterion
on the histograms allows detecting the presence or absence of a
bird nest. In this paper, branches, twigs and streaks have the same
meaning, and are interchangeably mentioned depending on
context.
Fig. 5. Trunk and branch identification after image binarization. (a) Original images (b) Images after binarization. (c) Detected trunks (d) Detected branches.
248
X. Wu et al. / Pattern Recognition 51 (2016) 242–254
4. Automatic bird nest detection for OCS images
4.1. Pre-Processing: Image binarization
Due to the convenience of non-contact on-board OCS inspection
system, the portable device with high resolution cameras is mounted
inside the driver's cabin instead of mounted on the roof of the train.
Therefore, the dusts on the train windows are easily captured by the
high-resolution cameras. In addition, the OCS images are always suf-
fering from noises, uneven lighting and compression artifacts. The
brightness distribution of various positions in an OCS image may vary
because of the condition of catenary and the effect of lighting envir-
onment. These reasons make noise a crucial factor affecting the
performance.
Binarization is the initial and a key step for automatic detection
of bird nests, since it is the base for successful localization and
recognition of bird nests. Its purpose is to distinguish the overhead
catenary areas from the background areas and remove noises. In
order to avoid the disadvantage of global threshold methods,
which use a single threshold value to classify image pixels into
objects and background classes, we use local adaptive binarization.
It deploys multiple values selected according to the local area
information. An OCS image is divided into multiple blocks
according to an n n sliding window, and for each block, the
average value of all pixels in the sliding window is selected as the
threshold.
In addition, we adopt two granularities of block size for the sliding
window to detect the main structures and small details, respectively.
The one with a larger size (e.g., 400 400) is to retain the shape of
relatively large objects, such as the supporter and the portal structure,
but ignore small details. On the contrary, the other one with a smaller
size (e.g., 10 10) is to keep details and thin edge information, such as
the streaks of bird nests. These two binarized images are then com-
bined by a union operation to form the final image. Their combination
perfectly contains the complete content of the overhead catenary and
the details such as lines and strokes of bird nests, which improves the
quality of catenary regions and preserves stroke connectivity by
removing isolated pixels. A post-processing technique is used to
eliminate noise pixels. Small connected regions with connected
component labeling [30] are treated as noises and then removed. The
original images and the effects after adaptive binarization are shown
in Fig. 5(a and b), respectively. After this step, each image is converted
from a gray scale to a binary image IB with reduced noises and shar-
pened objects.
4.2. Trunk and branch identification
After image binarization, the next step is to identify the trunk
and branch regions, which are the potential foundation of a bird
nest built and the twigs of a bird nest, respectively. The trunk in
the OCS images refers to the large regions such as the supporter
and the portal structure of the overhead catenary system. Mean-
while, the branch denotes thin lines and small details.
To identify the trunk and branch, Canny edge detector [31] is
performed on the binarized image to detect the edges, forming the
edge image IE, which makes the identification of the overhead contact
lines, the poles, and the streaks of bird nests easier. The main part of
the catenary and the portal structure is detected by morphological
opening operation with rectangle kernel of n n, which tends to
enlarge small holes, remove small objects, and separate objects. The
pixels between two edges whose distance is less than n/2 will be
filled. Otherwise, it will be kept unchanged. Finally, the trunk IT can be
achieved by combining the image IE after the aforementioned opera-
tions with the binarized image IB by a logical union operation. That is,
the intersection of the detected main part and the binarized image
will form the trunk regions, which is shown in Fig. 5(c). From this
figure, we can see that the main part of the supporter and the portal
structure of OCS can be correctly detected, which are the base where
bird nests will be built up. Once the trunks are extracted, the left part
is the thin lines and the twigs. The binarized image IB is subtracted
with the trunk image IT, which forms the branches and the thin lines.
The detected branch lines are shown in Fig. 5(d). We can see that the
thin structures, including the streaks of bird nests, are well identified.
4.3. Hovering point detection
Once the catenary structure is identified, the following step is
devoted to locate possible regions of a bird nest. The regions near the
main truck are potential areas that the bird nests will be constructed.
In addition, since the twigs of a bird nest are hovering in the air, the
hovering points for each twig are first identified to check the existence
of twigs, that is, the end of a twig. Ossification is undergone for the
image after the aforementioned steps. Therefore, the width of each
twig corresponds to one pixel. The next step is to find connected
components. The points satisfying the following two conditions,
connectivity and end point, are treated as hovering points. First, the
point itself is on the twig. It is connected with other points and it is
not an isolated point. Second, this point is at the end of a twig or a
line. In other words, the number of connected points among its
8 neighbors is just one. The condition of a hovering point is defined as
follows:
ðaÞ: Connectivity :
PCðx0 þm; y0 þnÞ ¼ 2
ð1Þ
m ¼ 1
n ¼ 1
ðbÞ: End po int : PCðx0; y0Þ ¼ 1
ð2Þ
where (x0,y0) is the location of a point and PC() is the number of
connected points in its 8 neighbors. For a hovering point, it has only
one connected point, such as point t in Fig. 6. On the contrary, a non-
hovering point has at least 2 connected points, such as points p and q
in Fig. 6. An example of the detected hovering points is shown in
Fig. 7, which are labeled in red color. If there are several hovering
points in a sliding window, it is a potential region having a bird nest.
However, it does not mean that it must be a bird nest in this region,
such as the insulator regions.
X1
X1
4.4. Twig streak extraction using Hough transform
Once the hovering points have been identified, we will extract
the twigs of a bird nest, which can be roughly represented as a set
of straight lines with differentiated lengths. Due to imperfections
in either the noises or the edge detector, there may be missing
Fig. 6. Hovering point detection. A hovering point (e.g., point t) has only one
connected point, while other points (e.g., points p and q) have more than two
connected points.
X. Wu et al. / Pattern Recognition 51 (2016) 242–254
249
points on the desired curves, as well as spatial deviations between
the ideal line and the noisy edge points obtained from the edge
detector. The image may contain multiple edge fragments corre-
sponding to a single whole streak. The classic Hough transform
[32] is famous for its capability with the identification of regular
curves in an image, such as lines, circles, etc. The main advantage
of the Hough transform is that it is tolerant to gaps in feature
boundary descriptions and is relatively unaffected by image noise.
To increase algorithm efficiency and to reduce computational
burden, the straight line search is limited inside the potential
Fig. 7. Detected hovering points of a bird nest candidate, which are labeled in red
dots. (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)
rectangular regions of the OCS image. Therefore, identifying twigs
is converted to the extraction of straight lines using Hough
transform, which is illustrated in Fig. 8. We can see that most of
the streaks are correctly identified although a small amount of
curves and short lines are ignored. When the straight lines are
detected, we can learn their patterns to check if there exists a bird
nest in the OCS image.
4.5. Pattern learning with HOS and HLS
Based on the properties of a bird nest, a bird nest is composed of a
bunch of short or long twigs with unordered, non-parallel organiza-
tion. In other words, the groups of streaks in images should be with
distributed and diverse orientations, at the same time, with incon-
sistent lengths. On the contrary, for regions such as insulators, the
streaks usually demonstrate uniform and parallel orientations and
they generally have equal lengths, as shown in Fig. 8. To represent the
orientation and length patterns of streaks in a bird nest, we construct
Histogram of Orientation of Streaks (HOS) and Histogram of Length of
Streaks (HLS), respectively, and then learn the patterns using SVM for
discriminative classification.
The histogram of orientations of streaks (HOS) can be constructed
by quantizing the angles of the detected streaks and the horizontal
axis, and then accumulating the angles for all streaks in the region.
The angle is in the range of 0°–180°. We quantize the angles into 18
bins with a step size of 10° from 0° to 180° to form a histogram. The
HOS histogram is constructed by counting the number of streaks at a
particular range of angles. The diverse distribution of HOS potentially
stands for the orientation of a bird nest while a uniform distribution
represents other cases such as insulators.
Similarly, the histogram of lengths of streaks (HLS) can be con-
structed by quantizing the lengths of the detected streaks, and then
accumulating the lengths for all streaks in the region. We quantize the
lengths equally into 13 bins with a step size of 5 pixels to form a
Fig. 8. The detected streaks in each region using Hough transform. (a) the original images, (b) the located regions, (c) the thin branches, (d) the detected lines using Hough
transform are labeled with different colors, (e) the final detected streaks. We can see that most of the streaks can be correctly detected.