PresentationPDF Available

Object Detection in Color Image

Authors:

Abstract

This is a presentation of the same-name report presented at the PRIP'2019 conference
SPIIRAS 1
Object Detection in Color Image
M.V. Kharinov, A.N.Buslavsky
khar@iias.spb.su
SPIIRAS
E
Subject: A model for preprocessing of images, audio-signals etc., where
preprocessing is treated as the generating of a set of representations
that might be used instead of a priori data.
In line with mainstream:
▪ practice: «PPAML» DARPA project 2013-2017,
theory: «Salient region detection» based on Ward’s clustering.
Key features of the model:
Supports the creation of software to detect object in image of any content
Provides: 1) Detection of the hierarchy of all objects in the image
2) Ordering of the objects by heterogeneity parameter that is
a quantitative measure of object saliency.
3) Online thresholding of objects by heterogeneity or area.
Object Detection in Color Image
M.V. Kharinov, A.N.Buslavsky
khar@iias.spb.su
SPIIRAS
SPIIRAS 2
Visual Model of All Object Detection by Pixel Clustering
262144
items
Image
Approximations
Dependence of the approximation
error on the number of clusters
E
g
Hierarchical sequence of quasi-
optimal image approximations
2
3NE
g
0
g
g
E
EE
E
gg
g
d
d
,
2
11
is standard deviation
Red line corresponds to optimal
approximations
Black line corresponds to
quasi-optimal approximations
is heteroge-
neity parameter
g0 is optimal cluster number
SPIIRAS 3
Computational Model of Pixel Clustering for Color Image
All calculations are performed in terms of the algebraic network formed by the
Sleator-Tarjan dynamic trees (acyclic graphs) and cycles (cyclic graphs).
The reversible computations in generalized form are supported.
Principal scheme of algebraic network construction
Merging of trees Merging of cycles
Resultant kernel network for an image of 4th pixels
Kernel tree Kernel cycle
SPIIRAS 4
Data Structure of Algebraic Network
Static (for storing&transmission) Dynamic (for processing&optimization)
Kernel
network
signal
R
G
B
Kernel tree
Kernel cycle
Kernel tree
Kernel cycle
R
G
B
Tree
Cycle
Cycle
Tree
Cycle
Tree
Tree
Cycle
Dynamic metadata for acce-
leration of computatons and
optimization of kernel network
SPIIRAS 5
Online Object Thresholding by Heterogeniety or Area
Image Object rating map
5 intensities, 14132 segments 13151 colors
Tuning parameters
1) Optimal cluster number g0=262144
2) Heterogeniety threshold |Edivide| = 1% (online parameter)
Averaged image
SPIIRAS 6
E Minimizing Operations
1) Merging:
2) Dividing:
3) Correcting:
0)2()1()21(
2
21
21
21
II
nn
nn
EEEE
merge
1
2
1
2
2
2
1111
nk
II
nk
II
E
correct
 
 
1
21
12
knn
knn
,0
21
IIIIE
correct
21
,, nnk
 
0)3,2(1)1(
mergedivide
EEE
, where
321
- numbers of pixels in clusters, - vectors of averaged intensities
21
,, III
Composite operation with pixel clusters
Merging/Structuring – After merging a pair of clusters for the resultant cluster, the
hierarchy of approximations is updated, so that the violations of convexity property
caused by merging are suppressed.
The Merging/Restructuring operation provides the transformation of any bottom-up
hierarchy of approximations into the hierarchy of approximations described by a
convex sequence of values
E
SPIIRAS 7
Basic Methods
1) Conventional Ward’s pixel clustering
and Mumford-Shah segmentation model
2) SI (Segmentation Improvement)
and ASI (Advanced Segmentation
Improvement) methods
3) K-meanless clustering/segmentation
methods
min
merge
E
0min
correct
E
dividemerge
EE
maxmin
Elementary
Composite
1) Mumford-Shah +SI+Ward ̶ conventional segmentation improvement
3) Merging/Structuring +ASI+Ward’s ̶improvement of any bottoom-up hierarchical pixel
clustering or image segmentation
4) Local Ward’s +ASI+Ward’s ̶pixel clustering, described by a convex sequence of
values
E
SPIIRAS 8
Conventional Processing Improvement
SPIIRAS 9
The complexity of modern images has dramatically increased and
continues to grow
but only the resolution increased. The number of objects
is growing insignificantly, so far.
Standard deviation doesn’t meet the visual perception
but for comparison of image with its own approximation it is quite adequate.
The calculating of optimal approximations is an NP-hard problem
but it doesn't matter for non-strict solutions.
The K-means method is a method that tends to minimize the MSE
but K-means is for arithmometers. This method use the truncated criterium and produces
false minima. To be replaced by the advanced method K-meanless (S.D.Dvoenko).
Practical image processing requires a priori data about objects
but the objects of interest are known to the programmer
using the target toolbox to create the intelligent software.
Misunderstood stereotypes in image processing
SPIIRAS 10
Conclusions
A model of a hierarchy of objects in a color image is designed to
provide a programmer-engineer with the program toolbox for creation
of intelligent machine vision software.
The model combines an accessible description of the objects in the
image and the equivalent computational data structure of the dynamic
network that is unfamiliar to most programmers so far.
So, for practical implementation of the model, it is useful firstly to
implement its elementary methods in modern high-level languages,
where the apparatus of Sleator-Taryan dynamic trees is intensively
introduced.
In line with the easy implementation of the model is the implementa-
tion of the K-meanless method instead of the outdated K-means in the
computing environment MatLab and/or others.
To exclude the reproduction of errors due to incorrect stereotypes, it
makes sense to slightly tweak the special educational courses related
to the use of cluster analysis in image processing.
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.