IRJET - A Review on Gradient Histograms for Texture Enhancement and Object Detection

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 03 | Mar 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 195
A Review on Gradient Histograms for Texture Enhancement and Object
Detection
Divyashree N.1, Dr. K. N. Pushpalatha2
1MTech DEC, Department of ECE, DSCE, Bangalore, India
2Associate Professor, Department of ECE, DSCE, Bangalore, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - Deblurring is the process of removing blurred artefacts from images, such as blur caused by defocusing of camera or
due to motion of the subject. Image priors such as non-local priors of image gradient are used, which play an important role in
image deblurring methods. Both local and non-local image priors improve noise and ringing artifacts while attenuating fine
textures. To solve this problem, the algorithm basedonGradientHistogramPreservation(GHP)isused. CombiningGHPmodelwith
non-local sparse prior constraint, the global and non-local sparse constraint will synthesize rich textures resulting in natural
image. For the resulting image, feature extraction using the Histogram Oriented Gradient (HOG) algorithm is used for object
detection. Due to their unpredictable appearance and the wide range of poses they can take, identifying objects in images is a
challenging task. The problem of object detection is the feature sets that demonstratelocallydefinedHistogramOrientedGradient
(HOG) descriptors and also provides excellent performance relative to other current feature sets. HOG is stable, scalable, and
effective in the extraction of features that operate on gradients between neighboring pixels. Using the Support Vector Machine
(SVM), the extracted function is then classified.
Key Words: Gradient Histogram Preservation (GHP), Non- locally Centralised Sparse Representation (NCSR), Histogram
Oriented Gradients (HOG), Support Vector Machines (SVM).
1. INTRODUCTION
Image processing improves the pictorial information and also employs methods capable of enhancing the image information
for human interpretation and analysis. It involves filtering of the noisy image and results in a filtered natural image which is
better than the noisy image. Image gets corrupted when noise is addedorwhen bluroccurs.Blurringoftheimageoccursdueto
camera shake, movement of the subject, blurred background or mist and fog. Blurring is a type of ideal image bandwidth
reduction due to imperfect image forming process. Blur can also be introduced by improperly focused lens or atmospheric
turbulence. In order to remove the blur, many image deblurring techniques try to reverse the degraded image in order to
recover the true image. Restoration often displays ringing artifacts and recovers missing components of the frequency.
Deblurring is an iterative process that considers the differentparametersfor eachiteration.Thisprocesswill continueuntil the
image received is based on the range of information which seems to be complete natural image.
Different image deblurring algorithms focus on developing appropriate regularizations to prevent the restored image
approaching the sharp latent image. Such concepts of regularization are based on priors of natural image, such as gradient
priors, non-local priors, and sparse priors. Priors are the image information that is used to define the natural image
characteristics. To differentiate restored images from unnatural images, another classical sparse or non-local prior model is
created. Nonetheless, most of the current image deblurring methods that leverage such regularizations caneliminatetextures
of mid-frequency while reducing image deblurring ringing artefacts and noise. Distinctions between priors result in the
depiction of different natural image characteristics, and thus should work moreeffectivelyincombiningdifferentimagepriors
on image deblurring.
Though the image is restored, there exists some variations in the intensitypixelscalledgradientsandcanbeestimatedthrough
graphical representation called histograms, in which the number of intensity pixels can be calculated. Also, fine textures are
attenuated during image deblurring. We use the gradient histogram preservation(GHP)model toaddressthisissue,according
to the intuitive thinking that the gradient of a well-reconstructed image should be the same asthatoftheoriginal image.Inthis
paper, we develop a texture-enhanced image deblurring algorithm (Gradient Histogram Preservation based Deblurring
algorithm, GHPD) combining the GHP model with the non-locally centralized sparse representation (NCSR). The proposed
denoising approach based on GHP can well boost the image texture regions,whichare oftenover-smoothedbyotherdenoising
methods. A novel image denoising improved texture method is introduced, which retains the original image's gradient
histogram. A gradient histogram preservation algorithm is developed using histogram specification to ensure the gradient
histogram of the denoised image is similar to the reference histogram.
A new deblurring algorithm is introduced, in which we attempt to exploit theNCSRmodel'snon-local similarity.Becauseofthe
homogeneity and randomness in an image between texture regions, the NCSR model will work well to eliminate noise and

ringing artefacts resulting from deblurring, whereasGHP performsbetterontextureenhancement.Thesparse prioristhemost
common gradient and NCSR model is one of this type, as stated earlier, this model-based deblurring approach can eliminate
substantial mid-level textures considered to be noise or ringing artefacts. Image deblurring will inevitably result in some
ringing artefacts and noise. Due to the degradation of the observed image, the sparse representations by conventional models
may not be accurate enough for a faithful reconstruction of the original image. At the end, the image is exploited by means of
nonlocal self-similarity to obtain good estimation of the sparse coding coefficientsoftheoriginal image,andthencentralizethe
sparse coding coefficients of the observed image to those estimation.
Object detection and classification are major techniques that are used in automated systems such as identification of persons,
self-driving cars, face recognition for security applications, text identification, disaster management, medical imaging.
Histogram Oriented Gradient is one of the common feature detectionalgorithms.HOGisfamousfordetectingthe boundariesof
the objects through gradient variation and the ability to classify varied pixels using linear SVM model. Dalal and Triggs
developed HOG [5] that operated on an INRIA pedestrian dataset. From the data it had correctly identified humans. Gradient-
based approaches have been an important area of research, and implementations are constantly being built in with more
human accuracy. It employs descriptors of local features based on edge and gradienttodefine representationoftheindividual.
It is easily controllable, translation and rotation invariant and produces good detection accuracies.
1.1 Feature Extraction
Feature is any quantifiable parameter obtained from an object in order to identify, analyze, describe and classify for a specific
application. Features are descriptive points to be identified from an object. Attributes in image processingareusuallycorners,
edges of an image object, regions of interest (ROI), and ridges. For object identification to be proficient, proper features has to
be selected for identifying the objects. Features should be unique, distinct, measurable and should be able to capture relevant
information and organize in an accessible manner for further reference.
Part-based approaches classify characteristics by splitting image into multiple parts and matchingtheloadeddata set.Holistic
methods identify an object which is viewed on a sliding window. Feature detection has been a key processing phenomenon to
machine vision, deep learning, classification and text analysis. Some of the feature extraction algorithms are Harris corner
detection, SIFT (Scale Invariant Feature Transform), SURF (Speeded-Up Robust Features), FAST (Features from Accelerated
Segment), FREAK (Fast Retina Key point) and HOG (Histogram of Oriented Gradients).
2. Related Works
R. Fergus, B.Singh, A. Hertzmann, S. Roweis and W.T. Freeman[1] analyzed a method of removing the camera shake effects in
photographs. By applying natural image priors, advanced statistical techniques and an extensively exploited natural images
through heavy-tailed gradient algorithm, the possible results were obtained. These approaches were useful in reducing other
computational photography problems.
The above approach resulted in few image variations. In order to remove these variations, L. Rudin, S. Osher and S. Fatemi[2]
proposed a method of minimizing the overall image variation using gradient projection method. A limited number of
algorithms for optimization is provided to extract noise from images. The overall picture variance is reduced to subject the
restrictions that include noise statistics. It amounts in solving a partial differential equationthatdependsuponmultiple oftime
defined by the constraints. The total variation of the image is minimized with respect to constraints that are imposed using
Lagrange multipliers and hence resulted in a restored image.
W. Dong, L. Zhang, G. Shi and X. Li [3] developed a sparse or non-local model is developed to differentiate between restored
images and unnatural images. For processing the image patches, one set of databases are selected to characterize the local
sparse domain regions. Two adaptive regularizationsare introducedintosparserepresentationframework.Thefirstmethodis
the autoregressive models which adaptively selects the regularizations in the image local structures.Thesecondmethodisthe
image nonlocal self-similarity is introduced in another regularization term. Thesparsityregularizationparameteris estimated
for better image restoration performance. Both the proposed method achieves much better results in both PSNR and visual
perception.
To extract the corner feature points during object detection, Harris C. and Stephens. M [4] proposed Harris corner detection
which identifies relevant feature points by employing a local detecting window on the image. Corners of an image show
important information about the quality of theimage wherevarious edgedirectionsmeet.Cornersareimportantfeaturepoints
because translation and rotation of pixels are invariant. Moving a detection window and observing the change of intensity of
pixels reveals the presence of an edge or corner.

Lowe and David G. [5] proposed SIFT algorithm which is a local and appearance-based method of extraction of featureswhich
is invariant to image scaling and rotation. Invariance to certain conditions is emphasized to define a robust distinctive feature
for representing an object/image even with variations in lighting,noiseandviewpoint.SIFTapproachisa basicconceptusedin
advanced techniques. SIFT performs well with small database in near real time efficiency.
Bay H., Tuytelaars T. and Van Gool L. [6] developed SURF algorithm which detects ROI points by an efficient algorithm which
was developed by Bay for representation invariant to neighborhood points and correlation. Selection of feature points,
obtaining a unique and robust feature to depict neighborhood and applying fast Hessian detector on coordinated descriptor
vector for extracting the feature points are the processes through which the algorithm works. The uniqueness of SURF results
in fast computation in feature matching points.
Rosten E. and Drummond T. [7] developed a FAST algorithm which is a corner detection method that uses segmentation of an
image into circles of 16 pixels and pixel intensity variation can be identified acrossthe bordersofthecircle.Asegmented testis
accelerated as the rejection of corner points that are fast and simple based on referral methods of some specified pixel points.
Vandergheynst P., Ortiz R and Alahi A. [8] proposed FREAK algorithm which is a binary matching descriptorthatimprovesthe
efficiency of descriptor by using a set of non-overlapping concentric circular rings similar to retinal pattern. The Gaussian
patterns around a point of interest are evaluated and the speed of thisalgorithmis bettercomparedtoexistingalgorithmssuch
as SIFT and SURF.
Suard, F et al. [9] designed a method for pedestrian detection for automotive in night light through stereo infrared images in a
single frame-based method. They have analyzed shapes based on the pixels classified using an SVM classifier.
Kobayashi, T et al. [10] used Principal Component Analysis based HOG to obtain a score, to be classifiedusinglinear SVM.PCA-
HOG is preferred for its ability to reduces dimensions, cost and improves the detection performance.
3. HISTOGRAM ORIENTED GRADIENTS
HOG is a gradient based feature descriptor and most widely used object detectionmethodknown foritssuperioridentification
character based on gradients, classification of gradients based on frequency of occurrence (histogram) and selection ofobject
category through support vector machines. The main application of HOG is pedestrian detection and has further extension to
traffic sign detection, face detection, hand writing digit recognition, landmine detection, disaster management, biomedical
imaging, etc. The histogram method for the directed gradients is as follows:
• Normalization of the captured image is performed forprocessingofintensityvaluesextractedfromthreecolorplanes(R,G,B
planes).
• Gradient asset of each picture element relative to its neighbor points is calculated.
• Classification of gradient pairs along mutually perpendicular axes, x and y directions, into incidence segment frequencies
implies as bins. A single 9 element vector is produced for every cell of 8 x 8 dimension.
• Any illumination on the classification accuracy can be avoided by generating blocks through cascading cells over which
intensities are normalized.
• The blocks are taken across the detection window to result in a string of HOG descriptors.
• The descriptors are fed to an SVM classifier which analysesthedata withpredefined weights,studiesthe characteristicsofthe
representative features and classifies the image as human or non-human.
Fig -1: Overview of Histogram Oriented Gradient

The classification accuracy is defined using Detection Error Trade-off (DET)curves.Errorcurvesprovidea comparisonofmiss
rate with the False Positives Per Window (FPPW) classified by the algorithm. Miss rate is the fraction of pedestrian cases that
has been erroneously declared as non-pedestrians. False Positive per window is the parameter that describes the ratio of
number of pedestrians that has been classified.
3. SUPPORT VECTOR MACHINES
SVM (Support Vector Machine) is a study system which uses linear functions as assumptionspaceforhighdimensional feature
space. This arithmetic is already used in human face detection and identification successfully, and it still widely used in
character, sound and others identification. Usually, linear SVM classifiers are used in conjunction with the HOG features for
object detection. It is mainly used for classification and regression. ThelinearSVMisa representationofdata amountsinwhich
a nonlinear projection of data into a high-dimensional space where it is easier to separate the two categories of data of image.
The radial basis function kernel, or RBF kernel is also known as Gaussian kernel,isa popularkernel functionusedforvarietyof
applications. It is widely used in support vector machine classifier for classification of data.
4. CONCLUSION
In this paper, we proposed a gradient histogram preservation-based texture enhanced model for image deblurring. By
constraint, the gradient histogram of the deblurred image approaching that of the reference image, the GHPD method can
achieve a satisfactory deblurring results with enhanced fine scale textures. We also applied this method to deblur the blurred
images and resulted in visual quality of the deblurred images. Furthermore, the GHPD method not only can be used for image
deblurring, but also for image denoising and other image reconstruction tasks. The main contribution of this technique is to
amplify the most significant features that describe particularly an object detection. The improved HOG feature represents
image, which depicts the edge features of image and reduces the impact of illumination. The improved HOG method largely
reduces computation and cost, consumption of energy and power and accelerates speed of detection and efficiency.
TABLE 1: COMPARISON OF RELATED WORKS
METHOD FUNCTIONALITY ADVANTAGES DISADVANTAGES
[1] Heavy tailed
gradient algorithm
Analyses the camera shake
effects of the photographs
Reduces the additional
photography problems
Results in pixel variation in the
image.
[2] Gradient
projection method
Minimizes the overall image
variation and picture
variances.
Total image variation of the
image is reduced and resulted in
a restored image.
Failed to differentiate between
natural and restored image.
[3] Adaptive sparse
representation
selection
Analyzed the difference
between natural and
restored image.
Adaptively selects the image
regularizations in the local
sparse domain which resulted in
both PSNR and visual perception.
Failed to detect the corners or
the edge feature points of the
image.
[4] Harris corner
detection algorithm
Identifies the local corner
feature points using
detecting window.
Shows theimportantinformation
of the image where various edge
directions meet.
Resulted in the variation of the
intensity pixels of the image.
[5] SIFT (Scale
Invariant Feature
Transform)
Extracts the image scaling
and rotation features
Performs well in small database
in near real time efficiency.
Varies in certain conditions like
noise and viewpointoftheimage.
[6] SURF (Speeded-
Up Robust Features)
Detects the ROI points by
fast Hessian detector.
Results in unique and robust
feature extraction of the image.
Variation in the neighborhood
points and correlation.
[7] FAST (Features
from Accelerated
Segment)
Corner detection of the
segmented image into
circles of 16 pixels.
Variation of the intensity pixel
across the borders of the image.
Rejects the corner points thatare
simple basedondifferent referral
methods.

[8] FREAK (Fast
Retina Key Point)
Uses binary descriptor by
using a set of non-
overlapping concentric
circular rings similar to
retinal pattern.
Provides good speed compared
to SIFT and SURF and improves
the efficiency.
Other featuredescriptorpatterns
result in mismatching of the
varied intensity pixels.
[9] Single frame-
based method
Pedestrian detection in
night light.
Analyses the shapes basedonthe
pixels classified using an SVM
classifier
Resulted in detecting only few
pedestrians due to lack of
brightness in an image.
[10] Principal
Component Analysis
based HOG
Analysis of object detection
using linear SVM
Cost efficiency and improves the
performance through PSNR and
RMSE.
Reduces the dimensions of the
object in an image.
REFERENCES
[1] R. Fergus, B. Singh, A. Hertzmann, S. Roweis, and W. T. Freeman, "Removing camera shake from a single photograph," ACM
Trans. Graph., 2006, vol. 25, no. 3, pp. 787-794.
[2] L. Rudin, S. Osher, and E. Fatemi, "Nonlinear total variation-based noise removal algorithms," Physica D: Nonlinear
Phenomena, 1992, vol. 60, no. 1, pp. 259-268.
[3] W. Dong, L. Zhang, G. Shi and X. Wu, "Image deblurring and super resolution by adaptive sparse domain selection.
[4] Harris, C., & Stephens. M, “A combined corner and edge detector”. In Alvey vision conference, Vol. 15, No. 50, 1988, pp. 10-
5244.
[5] Lowe, David G. "Object recognition from local scale-invariant features, " Proceedings of the International Conference on
Computer Vision, vol. 2, 1999, pp. 1150–1157.
[6] Bay, H., Tuytelaars, T., & Van Gool, L. “Surf: Speeded up robust features,” In European conference on computer vision
Springer, Berlin, Heidelberg, 2006, pp. 404-417.
[7] Rosten, E., & Drummond, T. “Machine learning for high-speed corner detection,” In European conference on computer
vision, Springer, Berlin, Heidelberg, 2006, pp. 430-443.
[8] Vandergheynst, P., Ortiz, R., & Alahi. A, “Freak: Fast retina key point,” In 2012 IEEE Conference on Computer Vision and
Pattern Recognition, IEEE. 2012, pp. 510-517.
[9] Suard, F., Rakotomamonjy, A., Bensrhair, A., & Broggi, A. “Pedestrian detection using infrared images and histograms of
oriented gradients,” In Intelligent Vehicles Symposium, IEEE, 2006, pp. 206-212.
[10] Kobayashi, T., Hidaka, A., & Kurita, T. “Selection of histograms of oriented gradients features for pedestrian detection,” In
International conference on neural information processing Springer, Berlin, Heidelberg, 2007, pp. 598-607.

IRJET - A Review on Gradient Histograms for Texture Enhancement and Object Detection

More Related Content

What's hot (19)

Similar to IRJET - A Review on Gradient Histograms for Texture Enhancement and Object Detection (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET - A Review on Gradient Histograms for Texture Enhancement and Object Detection