Machine Learning Approach for Object Detection - A Survey Approach

Description

Vol. 8 No. 6 September 2010 International Journal of Computer Science and Information Security

Shared by: ijcsis

Stats

views:: 600
posted:: 11/2/2010
language:: English
pages:: 5

Public Domain

Document Sample

(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 7, October 2010

Machine Learning Approach for Object
Detection - A Survey Approach
N.V. Balaji Dr. M. Punithavalli
Department of Computer Science, Karpagam Department of Computer Science, Sri Ramakrishna
University, Arts College for Women,
Coimbatore, India. Coimbatore, India.

Abstract---Object detection is a computer technology related to recently due to emerging applications which are not only
computer vision and image processing to determine whether or challenging but also computationally more demanding. These
not the specified object is present, and, if present, determine the applications include data mining, document classification,
locations and sizes of each object. Depending on the machine financial forecasting, organizing and retrieval of multimedia
learning algorithms, we can divide all object detection methods as databases, and biometrics and also the other fields where the
Generative Methods and Discriminative Methods. The concept of need of the image detection is high.
object detection is being an active area of research and it is
rapidly emerging since it is used in many areas of computer
vision, including image retrieval and video surveillance. This
paper presents a general survey which reviews the various
techniques for object detection and brings out the main outline of
object detection. The concepts of image detection are discussed in
detail along with examples and description. The most common &
significant algorithms for object detection are further discussed.
In this work an overview of the existing methodologies and
proposed techniques for object detection with future ideas for the
enhancement are discussed.
Keywords---Object Detection, Support Vector Machine, Neural
Networks, Machine Learning.

I. INTRODUCTION Figure 1. Description for the Image Detection
Extracting a feature vector of a given object and object
detection using the feature vector using pattern matching
The recognition problem is being posed as a classification
technique is the main goal for object detection [2]. Object
task, where the classes are either defined by the system
detection is to determine whether or not the object is present,
designer or are learned based on the similarity of patterns.
and, if present, determine the locations and sizes of each
Interest in the area of object detection has been renewed
object.
recently due to emerging applications which are not only
The most common approaches are image feature challenging but also computationally more demanding. These
extraction, feature transformation and machine learning where applications include data mining, document classification,
image feature extraction is to extract information about objects financial forecasting, organizing and retrieval of multimedia
from raw images. databases, and biometrics and also the other fields where the
need of the image detection is high.
Classification of patterns, object identification and its
description, are important tribulations to be concentrated upon
II. LITERATURE SURVEY
in a variety of engineering and scientific disciplines such as
biology, psychology, medicine, marketing, computer vision, Extraction of a reliable feature and improvement of the
artificial intelligence, and other remote sensing. Watanabe [1] classification accuracy have been among the main tasks in
defines a pattern as opposite of a chaos, that is, it is an entity, digital image processing. Finding the minimum number of
vaguely defined and that could be given a name. For instance, feature vectors, which represent observations with reduced
a pattern could be a fingerprint image, a handwritten cursive dimensionality without sacrificing the discriminating power of
word, a human face, or a speech signal. Given a pattern, the pattern classes, along with finding specific feature vectors, has
object detection may consist of one of the following two tasks been one of the most important problems in the field of pattern
[2] either the supervised classification in which the input analysis.
pattern is identified as a member of a predefined class or the In the last few years, the problem of recognizing object
unsupervised classification, which the pattern is assigned to a classes received growing attention in both variants of whole
previously unknown class. image classification and object localization. The majority of
The recognition problem is being posed as a classification existing methods use local image patches as basic features [3].
task, where the classes are either defined by the system Although these work well for some object classes such as
designer or are learned based on the similarity of patterns. motor-bikes and cars, other classes are defined by their shape
Interest in the area of object detection has been renewed and therefore better represented by contour features.

67 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 7, October 2010

In many real world applications such as pattern to carry out the generative task The problem of sophistication
recognition, data mining, and time-series prediction, we often the construction of a Bayesian network is known to be NP-
confront difficult situations where a complete set of training Hard, and therefore it is restricted to the structure of the
sample is not given when constructing a system. In face concluding network to a known form of arrangement to gain
recognition, for example, since human faces have large tractability.
variations due to expressions, lighting conditions, makeup,
hairstyles, and so forth, it is hard to consider all variations of The initial phase is to mine feature information from the
face in advance. object. Schneider man has discussed by using three level
wavelet transform to convert the contribution image to spatial
In many cases, training samples are provided only when a occurrence in sequence. One then constructs a set of
system misclassifies objects; hence the system is learned histograms in both position and intensity. The concentration
online to improve the classification performance. This type of values of each wavelet layer need be quantized to fit into an
learning is called incremental learning or continuous learning, inadequate number of bins. One difficulty encountered in the
and it has recently received a great attention in many practical premature execution of this method was the lacking of high
applications. power regularity information in the objects. With a linear
quantization scheme the higher energy bins had primarily
In pattern recognition and data mining, input data often singleton values, this leads to a problem when a prior is
have a large set of attribute. Hence, the informative input introduced to the bin, as the actual count values are lost in the
variables (features) are first extracted before the classification introduced prior. To extract this exponential quantization
is carried out. This means that when constructing an adaptive technique was employed to spread the power evenly between
classification system, we should consider two types of all the bin levels.
incremental learning: one is the incremental feature extraction,
and the other is incremental learning classifiers. D. Cluster-Based Object Detection
A. A hybrid object detection technique The cluster based object detection was proposed by Rikert,
Jones, and Viola [8]. In this methodology, the information
As discussed by M. Paul et. al., in [9] the adaptive about the object is learned and used for classification. The
background modeling based object detection techniques are objects are transformed and then build a mixture of Gaussian
widely used in machine vision applications for handling the model. The transformation is done based on the result of k-
challenges of real-world multimodal background. But they are means clustering applied to the transformed object. In the
forced to detailed environment due to relying on environment initial pace the object is distorted using a multi-directional
precise parameters, and their performances also alter across steer able pyramid. The result of the pyramid is then compiled
dissimilar operating speeds. The basic background calculation
into a succession of quality vectors self-possessed of the
is not appropriate for real applications due to manual foremost coat deposit pixel, and the pixels from higher in
background initialization prerequisite and its incapability to pyramid resized without interruption. For reasonably sized
switch cyclical multimodal background. It shows better patches this quickly becomes intractable.
firmness across different operating speeds and can better
abolish noise, shadow, and trailing effect than adaptive E. Rapid Object Detection using a Boosted Cascade of
techniques as no model adaptability or environment related Simple Features
parameters are involved. The hybrids object detection Paul Viola et. al., describe in [11], as a machine learning
technique for incorporating the strengths of both approaches. approach for object detection which is capable of processing
In this technique, Gaussian mixture models is used for images tremendously rapid and achieving high detection rates.
maintaining an adaptive background model and both This work is illustrious by three key contributions. The initial
probabilistic and basic subtraction decisions are utilized for is the prologue of an original object representation called the
scheming reasonably priced neighbor hood statistics for integral object which allows the features used by the detector
guiding the final object detection decision. to be computed very quickly. The author developed a learning
B. Moving Object Detection Algorithm algorithm, based on Ada Boost, which selects a small number
of critical visual features from a superior set and yields
Zhan Chaohui et. al., projected in [10], the first point in
enormously efficient classifiers [12]. The third contribution is
moving object detection algorithm is the block-based motion
a method for combining increasingly more complex classifiers
assessment is used to attain the common motion vectors, the
in a “cascade” which allows background regions of the object
vectors for every block, where the central pixel of the block is
to be quickly discarded while spending more calculation on
considered as the enter crucial point. These motion vectors are
showing potential object-like regions. The flow can be viewed
used to sense the border line blocks, which contain the border
as an object specific focus of concentration mechanism which
of the object. Presently on, the linear exclamation is used to
dissimilar to preceding approaches that provides statistical
make the coarse motion field an impenetrable motion field, by
guarantees that superfluous regions are improbable to contain
this way to eliminate the chunk artifacts. This possession can
the object of interest.
also be used to sense whether the motion field is uninterrupted
or not. This sophisticated impenetrable motion field is used to F. Template Matching Methods
define detail limitations in each boundary block. Thus the Huang T.S et. al., described the template matching
moving object is detected and coded. methods that uses standard patterns of objects and the object
C. Restricted Bayesian networks parts to portray the object globally or as diverse parts.
Correlations get struck between the input image and patterns
This approach presented by Schneiderman et. al., in [4, 5,
subsequently computed for detection. Gavrila [16] proposed
6 and 7] attempts to study the structure of a Bayesian network
an object detection scheme that segments forefront regions

68 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 7, October 2010

and extracts the boundary. Then the algorithm searches for images. They define a reduced set of regions that covers the
objects in the image by matching object features to a database image support and that spans various levels of resolution.
of templates. The matching is realized by computing the They are attractive for object detection as they enormously
average Chamfer detachment amid the template and the edge reduce the search space. In [23], several issues allied to the use
map of the target image area. Wren et al. [18] described of BPT for object detection are examined. Analysis in the
detailed on a top-down person detector based on template- compromise between computational complexity reduction and
matching. However, this approach requires field specific scene accuracy in accordance with the construction of binary tree
analysis. lead us to define two parts in BPT: one providing the accuracy
and the other representing the search space for the task of
G. Object Detection Using Hierarchical MRF and MAP object detection. In turn, it includes an analysis and
Estimation objectively compares various similarity measures for the tree
Qian R.J et. al., projected this method in [15] which construction. This different similarity criterion should be used
presents a new scale, position and direction invariant approach for the part providing accuracy in the BPT and for the part
to object detection. The technique initially chooses defining the search space. Binary Partition Tree concentrates
concentration on regions in an object based on the region in a compact and structured representation of meaningful
detection consequence on the object. Within the attention regions that can be extracted from an image. They offer a
regions, the method then detects targets that combine template multi-scale representation of the image and define the
matching methods with feature-based methods via hierarchical translation invariant 2-connectivity rule among regions.
MRF and MAP estimation. Hierarchical MRF and MAP
inference supply a stretchy framework to integrate various K. Statistical Object Detection Using Local Regression
visual clues. The amalgamation of template corresponding and Kernels
feature detection helps to accomplish robustness against This novel approach was proposed by Hae Jong Seo and
multifaceted backgrounds and fractional occlusions in object Peyman Milanfar in [24] to the problem of detection of visual
detection. similarity between a template image and patches in the given
image. The method is based on the computation of the local
H. Object Detection and Localization using Local and kernel of the template, which measures the likeness of a pixel
Global Features to its surroundings. This kernel is then used as a descriptor
The work proposed by Kevin Murphy et. al., in [21] from which features are extracted and compared against
describes more advanced method of object detection and analogous features from the target image. Comparison of the
localization using local and global features of an image. features extracted is carried out using canonical correlations
Traditional approaches to object detection only look at the analysis. The overall algorithm yields a scalar resemblance
local pieces of the image, whether it can be within a sliding map (RM). This resemblance map indicates the statistical
window or the regions around an interest point detector. When likelihood of similarity between a given template and all target
this object of interest is small or the imaging conditions are patches in an image. Similar objects with high accuracy can be
otherwise unfavorable, such local pieces of the image can obtained by performing statistical analysis on the resulting
become indistinct. This ambiguity can be reduced by using resemblance map. This proposed method is robust to various
global features of the image – which we call as a “gist” of the challenging conditions such as partial occlusions and
scene. The object detection rates can be significantly improved illumination change.
by combining the local and global features of the image. This
method also results in large increase of speed as well since the L. Spatial Histogram based Object Detection
gist is much cheaper to compute than the local detectors. Hongming Zhang et. al., describes in [25], that feature
extraction plays a major role for object representation in an
I. Object Detection from HS/MS and Multi-Platform Remote Automatic object detection system. The spatial histogram
Sensing Imagery preserves the object texture and shape simultaneously as it
Bo Wu et.al, put forth a technique in [22] that integrates contains marginal distribution of image over local patches. In
biologically and geometrically inspired approaches to detect [25], methods of learning informative features for spatial
objects from hyperspectral and/or multispectral (HS/MS), histogram-based object detection were proposed. Fisher
multiscale, multiplatform imagery. First, dimensionality criterion is employed to measure the discriminability of the
reduction methods are studied and implemented for spatial histogram feature and calculates features correlations
hyperspectral dimensionality reduction. Then, a biologically using mutual information. An informative selection algorithm
stimulated method S-LEGION (Spatial-Locally Excitatory was proposed in order to construct compact feature sets for
Globally Inhibitory Oscillator Network) is developed for efficient classification. This informative selection algorithm
object detection on the multispectral and dimension reduced selects the uncorrelated and discriminative spatial histogram
hyperspectral data. This method provides rough object shapes. features and this proposed method is efficient in object
Geometrically inspired method, GAC (Geometric Active detection.
Contour), is employed for refining object boundary detection
on the high resolution imagery based upon the initial object M. Recursive Neural Networks for Object Detection
shapes provided by S-LEGION. M. Bianchini et. al., put forth an algorithm in [26], a new
recursive neural network model for object detection. This
J. Binary Partition Tree for Object Detection algorithm is capable of processing directed acyclic graphs
This proposal suggested by V. Vilaplana et. al., in [23], with labeled edges, which address the problem of object
discusses the use of Binary Partition Tree (BPT) for object detection. The preliminary step in an object detection system
detection. BPTs are hierarchal region based representation of is the detection. The proposed method describes a graph-based

69 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 7, October 2010

representation of images that combines both spatial and visual learning or continuous learning. This problem is to introduce
features. The adjacency relationship between two Incremental Linear Discriminant Analysis as the feature
homogeneous regions after segmentation can be determined extraction technique to object detection and hence improving
by the edge between the two nodes. This edge label collects the classification performance to the great height.
information on their relative positions, whereas node labels
contain visual and geometric information on each region (area, The overall outcome of the proposed work is to implement
color, texture, etc). These graphs are then processed by the a variation in the existing feature extraction system LDA and
recursive model in order to determine the eventual presence to develop a new system ILDA which increases the
and the position of objects inside the image. The proposed classification performance to a great height. Also the system
system is general and can be employed to any object detection should take new samples as input online and learn them
systems, since it does not involve any prior knowledge on any quickly. As a result of this incremental learning process, the
particular problem. system will have a large set of samples learned and hence will
decrease the chance of misclassifying an object.
N. Object Detection Using a Shape Codebook
Object detection by Xiaodong Yu et. al., in [27], presents a IV. CONCLUSION
method for detecting categories of object in real world images. This paper attempts to provide a comprehensive survey of
The ultimate aim is to localize and recognize instances in the research on object detection and to provide some structural
training images of an object category. The main contribution categories for the methods described. When appropriately
of this work is a novel structure of the shape code-book for considered, it is been reported that, on the relative
object detection. The code book entry consists of two performance of methods in so doing, it needs the awareness
components: a shape codeword and a group of associated that there is a lack of uniformity in how methods are evaluated
vectors that specify the object centroids. The shape codeword and so it is reckless to overtly state that which methods indeed
is such that it can be easily extracted from most image object have the lowest error rates. As a substitution, it can be urged
categories. A geometrical relationship between the shape to the members of the community to expand and contribute to
codeword is stored by the associated vectors. The test sets and to report results on already available test sets. The
characteristics of a particular object category can be specified community needs to more seriously considered for systematic
by the geometrical relationship. performance evaluation. This would allow users and the
Triple-Adjacent-Segments (TAS), extracted from image researchers of the object detection algorithms to identify
edges is used as a shape codeword. Object detection is carried which ones are aggressive in which particular domain. It will
out in a probabilistic voting framework. This proposed method also prompt researchers to produce truly more effective object
has drastically lower complexity and requires noticeably less detection algorithms.
supervision in training. REFERENCES
O. Contour-based Object Detection in Range Images [1] S. Watanabe, Pattern Recognition: Human and Mechanical. New York:
Wiley, 2005.
This approach investigated by Stefan Stiene et. al., in [28], [2] Ilin, R. Kozma, R. Werbos, P.J, "Beyond Feedforward Models Trained
projects a novel object recognition approach based on range by Backpropagation: A Practical Training Tool for a More Efficient
images. Due to the insensitivity to illumination, range data is Universal Approximator", IEEE Transactions on Neural Networks, Vol.
most suited for reliable outline extraction. This determines 19, No. 6, June 2008.
(silhouette or contour descriptions) as good sources for object [3] Culp, M. Michailidis, G, "Graph-Based Semisupervised Learning ",
IEEE Transactions on Pattern Analysis and Machine Intelligence,
recognition. Based on a 3D laser scanner, contour extraction Volume: 30 Issue: 1, 2008.
performed using floor interpretation. Feature extraction is [4] H. Schneiderman, "A Statistical Approach to 3d Object Detection
done using a new fast Eigen-CSS method and a supervised Applied to Faces And Cars", 2000.
learning algorithm. It proposes a complete object recognition [5] H. Schneiderman. Learning statistical structure for object detection,
system. This recognition system was found to be tested 2003.
[6] H. Schneiderman. Feature-centric evaluation for efficient cascaded
successful on range images captured with the help of mobile
object detection, 2004.
robot. This results is compared with standard techniques i.e., [7] H. Schneider man. Learning a restricted Bayesian network for object
Geometric features, Border signature method, and the angular detection. 2004.
radial transformation. The Eigen-CSS method is found to be [8] P. Viola T. Rikert and M. Jones. A cluster-based statistical model for
more efficient than the best one by an order of magnitude in object detection.1999.
feature extraction time. [9] Haque, M. Murshed, M. Paul, M, “A hybrid object detection
technique from dynamic background using Gaussian mixture models”,
Gippsland Sch. of Inf. Technol., Monash Univ., Clayton, VIC,
III. FUTURE ENHANCEMENT Multimedia Signal Processing, 2008 IEEE 10th Workshop, Oct. 2008.
[10] Zhan Chaohui, Duan Xiaohui, Xu Shuoyu, Song Zheng and Luo Min,
The object detection methodologies are improving day by
“An Improved Moving Object Detection Algorithm Based on Frame
day as its need is hastily growing. Hence the techniques Difference and Edge Detection”, ICIG 2007. Fourth International
consider feature extraction where the Principal Component Conference, Aug. 2007.
Analysis and Linear Discriminant Analysis are the most [11] Paul Viola and Michael Jones, “Rapid Object Detection using a Boosted
common approaches available. In object detection systems, the Cascade of Simple Features”, IEEE 2001.
complete set of samples is not given at the time constructing [12] Yoav Freund and Robert E. Schapire, “A decision-theoretic
generalization of on-line learning and an application to boosting”, In
the system. Instead, more and more samples are added Computational Learning Theory: Eurocolt ’05, Springer-Verlag, 2005.
whenever the system misclassifies objects and hence the [13] Sharaf, R. Noureldin, A, "Sensor Integration for Satellite-Based
system is learned online to improve the classification Vehicular Navigation Using Neural Networks", IEEE Transactions on
performance. This type of learning is called incremental Neural Networks, March 2007.

70 http://sites.google.com/site/ijcsis/
ISSN 1947-5500
(IJCSIS) International Journal of Computer Science and Information Security,
Vol. 8, No. 7, October 2010

[14] Juyang Weng Yilu Zhang Wey-Shiuan Hwang , "Candid covariance-
free incremental principal component analysis", IEEE Transactions on
Pattern Analysis and Machine Intelligence, 2003.
[15] Qian R.J and Huang T.S, “Object detection using hierarchical MRF and
MAP estimation”, Vision and Pattern Recognition, 2007. Proceedings.
2007 IEEE Computer Society Conference.
[16] D.M. Gavrila and V. Philomin, “Real-time object detection for smart
vehicles”, IEEE Computer Society Conference on Computer Vision and
Pattern Recognition Workshops (CVPR), pp. 87–93, 1999.
[17] P. Viola and M. Jones, “Rapid object detection using a boosted cascade
of simple features”, In Proc. of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), vol. 1, pp. 511-518, 2001.
[18] C. R. Wren, A. Azarbayejani, T. Darrell, and A. Pentland, “PFinder:
real-time tracking of the human body”, IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 19, no. 7, pp. 780-785, 1997.
[19] M. A. Aizerman, E. M. Braverman, and L. I. Rozonoer, “Theoretical
foundations of the potential function method in pattern recognition
learning,” Automat. Remote Contr., vol. 25, pp. 917–936, 2004.
[20] L. I. Rozonoer, “The probability problem of pattern recognition learning
and the method of potential functions,” Automat. Remote Contr., vol.
25, 2004.
[21] Kevin Murphy, Antonio Torralba, Daniel Eaton, and William Freeman.
Object detection and localization using local and global features. 2007.
[22] Bo Wu, Yuan Zhou, Lin Yan, Jiangye Yuan, Ron Li, and DeLiang
Wang, “Object detection from HS/MS and multi-platform remote
sensing imagery by the integration of biologically and geometrically
inspired approaches,” ASPRS 2009 Annual Conference, Baltimore,
Maryland. March 9-13, 2009.
[23] V. Vilaplana, F. Marques and P. Salembier, “Binary Partition Tree for
object detection”, Image Processing, IEEE Transactions on Image
Processing, vol. 17, no. 11, pp. 2201-2216, 2008.
[24] Hae Jong Seo and Peyman Milanfar, “Using local regression kernels for
statistical object detection,” IEEE Transaction on Pattern Analysis and
Machine Intelligence 2008.
[25] Hongming Zhang, Wen Gao, Xilin Chen, and Debin Zhao, “Learning
informative features for spatial histogram-based object detection,” IEEE
2005. International Joint Conference on Neural Networks, Montreal,
Canada, July 31-August 04, 2005.
[26] M. Bianchini, M. Maggini, L. Sarti, and F. Scarselli, “Recursive neural
networks for object detection,” IEEE Transaction on Pattern Analysis
and Machine Intelligence 2004.
[27] Xiaodong Yu, Li Yi, Cornelia Fermuller, and David Doermann, “Object
detection using a shape code-book,” IEEE, 2007.
[28] Stefan Stiene, Kai Lingemann, Andreas Nuchter, and Joachim
Hertzberg, “Contour-based Object detection in range images,” IEEE
2006.
AUTHORS PROFILE

N.V. Balaji has obtained his Bachelor of Science in
Computer Science from Sri Ramasamy Naidu Memorial
College, Sattur in 1997 and Master of Science in
Computer Science from Dr. GRD College of Science in
1997. Now he is doing Ph.D., in Bharathiar University.
He commences more than nine years of experience in
teaching field moreover industrial experience in Cicada
Solutions, Bangalore. At present he is working as Asst.
Professor & Training Officer at Karpagam University. His research interests
are in the area of Image Processing and Networks. He presented number of
papers in reputed national and international journals and conferences.

Dr. M. Punithavalli received the Ph.D degree in
Computer Science from Alagappa University, Karaikudi
in May 2007. She is currently serving as the Adjunct
Professor in Computer Application Department, Sri
Ramakrishna Engineering College, Coimbatore. Her
research interest lies in the area of Data mining, Genetic
Algorithms and Image Processing. She has published
more than 10 Technical papers in International, National
Journals and conferences. She is Board of studies member various
universities and colleges. She is also reviewer in International Journals. She
has given many guest lecturers and acted as chairperson in conference.
Currently 10 students are doing Ph.D., under her supervision.

71 http://sites.google.com/site/ijcsis/
ISSN 1947-5500