Download Image Clustering For Feature Detection

Image Clustering For
Feature Extraction
Using K-mean Clustering and Viola & Jones
Object Detection technique
DATA MINING
Zeeshan Anjum EP 115051
BSSE VI
Muzammil Hussain Zaidi EP115043
Submitted to Dr. Tahseen Ahmed Jilani
Index
Zeeshan Anjum EP115051 |
Muzammil Hussain Zaidi EP
115043
Introduction of Image
Classification
1.
What is Classification?
The classification problem becomes very hard when there are many
parameters. There are so many different combinations of parameters that
techniques based on exhaustive searches of the parameter space are
computationally infeasible. Practical methods for classification always
involve a heuristic approach intended to find a “good-enough” solution to
the optimization problem.
Zeeshan Anjum EP115051 |
Muzammil Hussain Zaidi EP
115043
Abstract
There are various method to classify
images but most of the time we have
calculate those area features which
aren’t much useful, usually those are
very small in ratio to overall image. So
we have introduce a new technique
by which it would be easier for any
classifier to classify any image or
image sequence (real-time) in a
fractions of second.
First we blur the image to reduce
noise from image, because in next
step we are going to find k-mean
clusters where k is specified before;
using some predetermined value. It
must be dynamic to overcome color
variation problem. Next we are going
to use canny edge detection method
we can also use Sober, Robert cross
or Prewitt as well. The reason for
using clustering before edge detection
is to minimize small details such as
patch, hole, dirt etc. as our aim in this
project is to find the clear area where
an autonomous vehicle can drive with
the extremely lowest value of error.
When k-clusters are detected, it will
generate 2-D map of RGB values with
high contrast edges, and provide
flexibility to edge detection algorithm.
When proper edges are detected. We
pass that image to classifier at this
stage generated edges should be
scaled up to half the size of original
image and that image size should be
reduced to match both the image and
edge filter. We can only use edge
feature to predict feature such as
board, road, U-turn, crossing, square
etc. but some advance feature
required the actual image to merge
with these edges to boost the feature
edge pixel intensity.
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Data Mining
Data mining (the analysis step of the "Knowledge Discovery and Data Mining"
process, or KDD), an interdisciplinary subfield of computer science, is the
computational process of discovering patterns in large data sets involving
methods at the intersection of artificial intelligence, machine learning,
statistics, and database systems. The overall goal of the data mining process is
to extract information from a data set and transform it into an understandable
structure for further use. Aside from the raw analysis step, it involves database
and data management aspects, data pre-processing, model and inference
considerations, interestingness metrics, complexity considerations, postprocessing of discovered structures, visualization, and online updating.
The term is a buzzword, and is frequently misused to mean any form of largescale data or information processing (collection, extraction, warehousing,
analysis, and statistics) but is also generalized to any kind of computer decision
support system, including artificial intelligence, machine learning, and business
intelligence. In the proper use of the word, the key term is discovery[citation
needed], commonly defined as "detecting something new". Even the popular
book "Data mining: Practical machine learning tools and techniques with Java"
(which covers mostly machine learning material) was originally to be named
just "Practical machine learning", and the term "data mining" was only added
for marketing reasons. Often the more general terms "(large scale) data
analysis", or "analytics" – or when referring to actual methods, artificial
intelligence and machine learning – are more appropriate.
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Data Mining
The actual data mining task is the automatic or semi-automatic analysis of large
quantities of data to extract previously unknown interesting patterns such as
groups of data records (cluster analysis), unusual records (anomaly detection)
and dependencies (association rule mining). This usually involves using
database techniques such as spatial indices. These patterns can then be seen as
a kind of summary of the input data, and may be used in further analysis or, for
example, in machine learning and predictive analytics. For example, the data
mining step might identify multiple groups in the data, which can then be used
to obtain more accurate prediction results by a decision support system.
Neither the data collection, data preparation, nor result interpretation and
reporting are part of the data mining step, but do belong to the overall KDD
process as additional steps.
The related terms data dredging, data fishing, and data snooping refer to the
use of data mining methods to sample parts of a larger population data set that
are (or may be) too small for reliable statistical inferences to be made about the
validity of any patterns discovered. These methods can, however, be used in
creating new hypotheses to test against the larger data populations.
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Machine Learning
Machine learning, a branch of artificial intelligence, concerns the
construction and study of systems that can learn from data. For
example, a machine learning system could be trained on email
messages to learn to distinguish between spam and non-spam
messages. After learning, it can then be used to classify new email
messages into spam and non-spam folders.
The core of machine learning deals with representation and
generalization. Representation of data instances and functions
evaluated on these instances are part of all machine learning systems.
Generalization is the property that the system will perform well on
unseen data instances; the conditions under which this can be
guaranteed are a key object of study in the subfield of computational
learning theory.
“Field of study that gives computers the ability to learn without being explicitly
programmed”
1959, Arthur Samuel
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Data Mining and Machine Learning
These two terms are commonly confused, as they often employ the same methods and
overlap significantly. They can be roughly defined as follows:
Machine Learning focuses on prediction, based on known properties learned from the
training data.
Data mining focuses on the discovery of (previously) unknown properties in the data.
This is the analysis step of Knowledge Discovery in Databases.
The two areas overlap in many ways: data mining uses many machine learning methods,
but often with a slightly different goal in mind. On the other hand, machine learning also
employs data mining methods as "unsupervised learning" or as a preprocessing step to
improve learner accuracy. Much of the confusion between these two research
communities (which do often have separate conferences and separate journals, ECML
PKDD being a major exception) comes from the basic assumptions they work with: in
machine learning, performance is usually evaluated with respect to the ability to
reproduce known knowledge, while in Knowledge Discovery and Data Mining (KDD) the
key task is the discovery of previously unknown knowledge. Evaluated with respect to
known knowledge, an uninformed (unsupervised) method will easily be outperformed by
supervised methods, while in a typical KDD task, supervised methods cannot be used due
to the unavailability of training data.
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Cluster analysis
Cluster analysis or clustering is the task of grouping a set of objects in such a way that
objects in the same group (called a cluster) are more similar (in some sense or another) to
each other than to those in other groups (clusters). It is a main task of exploratory data
mining, and a common technique for statistical data analysis, used in many fields,
including machine learning, pattern recognition, image analysis, information retrieval, and
bioinformatics.
onlinecourses.science.psu.edu
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
K-Mean Clustering
k-means clustering is a method of vector quantization, originally
from signal processing, that is popular for cluster analysis in data
mining. k-means clustering aims to partition n observations into k
clusters in which each observation belongs to the cluster with the
nearest mean, serving as a prototype of the cluster. This results in a
partitioning of the data space into Voronoi cells.
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Steps of Feature
Classification
•
•
•
•
It is difficult to convolve image with large dimension so make a small
frame of 240x240 or 360x360, this size can be vary depends on the
computational power and accuracy requirement.
Now convolve image I with blur filter like Gaussian or LoG Laplacian of
Gaussian kernel. This method help us to reduce noise. But in our given
example we have implemented disk filter in matlab. Which can be used
as “fspecial('disk', radius) returns a circular averaging filter (pillbox)
within the square matrix of side 2*radius+1. The default radius is 5.”
It is best practice to use desired color space control video recording
device, in our example we convert RGB to LAB color space because the
reason for this conversion is to get lightness separately from color
opponent dimensions. It is defined as L, a, b. I suggest to use this
conversion otherwise it would produce noisy objects with features.
Since we have now converted values to a and b, we can now apply kmean algorithm, remember to use a good estimated value depends on
the environment factor.
–
–
–
•
High cluster value could result in detail segmentation of road like patches, dirt
etc.
Low cluster value could be result in minimum feature detection as in our
example, when we use 2 as the value of cluster it make a separate sky and
complete earth as 2 cluster, sometimes clouds and rest of the environment as
2 separate cluster.
Idea to choose the good value is depends on the view we are focusing, we
use 3-4 cluster to get rid out of the problem.
Now calculate k-mean using distance formula either Euclidean or city
block we use city block, because to get rid out of square root calculation
for each pixel pair. We limit this process to 4, this make sense because in
real-time at each second you have to calculate at least 10 frames, with
dynamic changing of environment. We are assuming that we‘ll get
satisfactory result from 4 time centroid shifting.
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Steps of Feature
Classification
–
•
•
•
•
Now we have we have color segmented image.
Next we are going to apply canny edge detection filter to get fine
tuned edges from segmented image.
Advantage of this technique is to reduce load of computation and
memory on onboard autonomous vehicle. Application includes
AGV,AAV etc.
Disadvantage is that this method is more error prune to high density
of normalized color in frame.
–
•
Then we reshape that array using reshape function, which takes an 1-D
Array with 2 parameters m, n such that B = reshape( A, m, n) where A is
1-D and satisfy size(A) = m*n , B is n-dimension matrix. This function has
other overloaded functions as well.
Solution: system should build in a way that it can switch this algorithm
b/w other with the constrain of color space. It is highly suggested that
fuzzy logic model should be used instead of Boolean logic for switching
operation to reduce frequent switching of segmentation methods.
Next the segmented image is passed to Classifier which used the
technique of Adaptive Boosting by combining weak classifiers into a
stronger one. This technique use semi automatic machine learning,
we have to provide Haar cascades of possible interested geometry.
The best technique to classify the image is to use Viola–Jones object
detection framework.
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Workflow of Classification
Input Image > Blur > K-Mean > Canny Edge Detection > Merge Image + Edges
Original
K-mean Clusters (6)
Canny Edge
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Example
Original
K-mean Clusters (6)
Canny Edge
a
K-mean Clusters (3)
b
K-mean Clusters (3)
c
It is clearly visible how perfectly this technique make
clusters of interesting patterns. Lane mark on image 3
cause problem because of highly color separation. In this
classification but in next stage (classification) it can be
easily removed using pattern detection.
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Code- Matlab
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
clear;
clc;
img = imread('02.jpg');
backup=img;
%blur this image out
H = fspecial('disk',3);
blurred=imfilter(img,H,'replicate',1);
img=blurred;
%doing main stuff
cform = makecform('srgb2lab');
lab = applycform(img,cform);
ab = double(lab(:,:,2:3));
nrows =size(ab,1);
ncols=size(ab,2);
ab = reshape(ab,nrows*ncols,2);
nColors=4;
[cluster_idx cluster_center] = kmeans(ab,nColors,'distance','cityblock',...
'Replicates',6);
pxl_lbl=reshape(cluster_idx,nrows,ncols);
ZX=edge(pxl_lbl,'canny');
%imshow(ZX,[])
segmented_images = cell(9,9);
%rgb_label = repmat(ZX,[1 1 1]);
%rgb_label = repmat(ZX,[0 0 0]);
rgb_label = repmat(ZX,[1 1 1]);
for k = 1:nColors
color = img;
color(rgb_label ~= k) =1;
segmented_images{k} = color;
end
%imshow(segmented_images{1})
%imshow(pxl_lbl,[]);
imshow(ZX,[])
%imshow(segmented_images{1})
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Classification using Viola and Jones
•
Feature extraction and feature evaluation.
– Rectangular features are used, with a new image
representation their calculation is very fast.
•
Features are extracted from sub windows of a sample
image.
The base size for a sub window is 24 by 24 pixels.
Each of the four feature types are scaled and shifted across all
possible combinations
In a 24 pixel by 24 pixel sub window there are ~160,000
possible features to be calculated.
•
Classifier training and feature selection using a slight
variation of a method called AdaBoost.
•
A combination of simple classifiers is very effective since
it is using multiple weak classifier to make a stronger
classifier.
Four basic types.
•
– They are easy to calculate.
– The white areas are subtracted from the black ones.
– A special representation of the sample called the integral
image makes feature extraction faster.
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Integral Image And Haar Cascade
These cascade starts with either a large (or small) window and
scans the image exhaustively (i.e. such as by dislocating the
window some points to the right, and going down at the end of
the line). When a scan finishes, it shrinks (or grows) this
window, repeating the process all over again.
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
AdaBoost
The classification scheme used by the Viola-Jones method is actually a cascade of boosted
classifiers. Each stage in the cascade is itself a strong classifier, in the sense it can obtain a
really high rejection rate by combining a series of weaker classifiers in some fashion.
A weak classifier is a classifier which can operate only marginally better than chance. This
means it is only slightly better than flipping a coin and deciding if there is something in the
image or not. Nevertheless, it is possible to build a strong classifier by combining the
decision of many weak classifiers into a single, weighted decision. This process of
combining several weak learners to form a strong learner is called boosting. Learning a
classifier like this can be performed, for example, using many of the variants of the
AdaBoost learning algorithm.
Haar Cascades link:
http://www.cs.utah.edu/~turcsans/DUC_files/
and C# code is from open source library.
opencv also used in DARPA Grand Challenge.
Reference:
Viola and Jones, "Rapid object detection using a boosted
cascade of simple features", Computer Vision and Pattern
Recognition, 2001
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Creating new Haar cascades for a unique/new object we have to required ~80% positive
images and ~20% negative images. This task is very easy but extremely time consuming
because a good classifier required approx. 2000 images of a feature which is sometime
not appropriate.
Cascade of Classifiers
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043
Reference and links
•
•
•
•
•
•
•
•
•
•
•
Fayyad, Usama; Piatetsky-Shapiro, Gregory; Smyth, Padhraic (1996). "From Data Mining
to Knowledge Discovery in Databases". Retrieved 17 December 2008.
"Data Mining Curriculum". ACM SIGKDD. 2006-04-30. Retrieved 2011-10-28. Clifton,
Christopher (2010). "Encyclopædia Britannica: Definition of Data Mining". Retrieved 201012-09.
Hastie, Trevor; Tibshirani, Robert; Friedman, Jerome (2009). "The Elements of Statistical
Learning: Data Mining, Inference, and Prediction“
See e.g. OKAIRP 2005 Fall Conference, Arizona State University, About.com: Datamining
Witten, Ian H.; Frank, Eibe; Hall, Mark A. (30 January 2011). Data Mining: Practical
Machine Learning Tools and Techniques (3 ed.). Elsevier. ISBN 978-0-12-374856-0.
Bouckaert, Remco R.; Frank, Eibe; Hall, Mark A.; Holmes, Geoffrey; Pfahringer, Bernhard;
Reutemann, Peter; Witten, Ian H. (2010). "WEKA Experiences with a Java open-source
project". Journal of Machine Learning Research 11: 2533–2541.
http://en.wikipedia.org/wiki/Machine_learning#Machine_learning_and_data_mining
http://en.wikipedia.org/wiki/K-means_clustering
http://reason.cs.uiuc.edu/jaesik/research/research-computer-vision/realtime-on-roadvehicle-detection/
Jaesik Choi , Realtime On-Road Vehicle Detection with Optical Flows and Haar-Like Feature
Detector, Computer Science Research and Tech Reports, University of Illinois at UrbanaChampaign, 2012.
onlinecourses.science.psu.edu
Zeeshan Anjum EP115051 | Muzammil Hussain Zaidi EP 115043