Вы находитесь на странице: 1из 4

High Quality Image Segmentation

Apurv Sirohi Abhishek Vishwakarma Ashish Ranjan


Computer Science and Engineering Computer Science and Engineering Computer Science and Engineering
PDPM IIITDM PDPM IIITDM PDPM IIITDM
Jabalpur Jabalpur Jabalpur
2017054 2017016 2017061

Abstract: In the context of image segmentation For the purpose of image segmentation some extra nodes
for various image processing applications, a high denoted as terminals are required. Each terminal
quality segmentation of images is mandatory. corresponds to one of the possible pixel labels. For binary
Assuming a Graphical representation of the segmentation we need two terminals, the source s and the
images, we followed min-cut/max-flow sink t. In a terminal graph, each pixel is connected not only
segmentation algorithm. We used an original to its 8 neighbors, but also to the terminal nodes s and t.
initialization of the graph cut algorithm based on Hence, we distinguish two kinds of edges: n-links
the segmentation result of Boykov-Kolmogorov (neighboring links) are edges between two pixel nodes
approach. We then proposed our improvement for and t-links (terminal links) are edges between a pixel node
image segmentation using Simple Linear Iterative and a terminal. Thus, the two-terminal graph is defined as:
Clustering (SLIC) algorithm that efficiently uses
the color differentiation formula. Both sets of
results have been automatically classified
according to a set of significant features. Results
are compared

Keywords—component, formatting, style, styling, insert (key


words)

I. INTRODUCTION
Image segmentation involves the identification of regions of
interest, which generally are an object or a part in a digital
image. Each region must maximize the homogeneity of its
pixel’s features (color, texture) and simultaneously
maximize the differences with neighboring regions;
moreover, each region must be spatially compact. In
general, image segmentation is considered an important
initial step in most image processing applications, which use
the segmentation information in order to perform upper
level tasks, such as object tracking or identification, and
scenes interpretation.

Graph theory is a very important aspect in implementing the


image segmentation. We have used two methods here, one
being the traditional max-flow/min-cut method that A two terminal graph representing a 3X3 image
implements Edmond-Karp algorithm for solving the graph
and finding the augmenting path through which the image is
cut. The second method involves the use of SLIC(Simple II. IMAGE SEGMENTATION USING MAX-FLOW
Linear Iterative Clustering) for the formation of image IMPLEMENTATION[1]
clusters, cie-de-2000 for calculation of the intensities and
the histogram plot to assign the value to the edges which is Min-Cut/Max-Flow
then solved using maxflow to find the augmenting path.
Grid graphs are typically employed in computer vision to
represent images, since the alignment of nodes in rows and An s/t cut C (hereafter referred to only by cut) on a
columns is a natural representation of the image pixels p∈P. graph separates the set of nodes V into two disjoint
A neighborhood system, N={{p,q}|p,q∈P}, has to be chosen subsets S⊂V and T⊂V, S∩T=∅, such that the
to establish the edges configuration connecting the different source s∈S and the sink t∈T. In our
pixel nodes. For our application, the second order application, S corresponds to the shw label and T to the
neighborhood system is considered, which implies that each bkg label. A cut C={S,T} is a subset of E containing all
pixel is connected to its 8 direct neighbors. edges {p,q} where p∈S and q∈T. The cost of a cut |C| is

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE


defined as the sum of the weights of the edges in C, network Gf=(V,Ef) to be the same network but with
i.e., |C|=Σ{p,q}∈Cw{p,q}. The minimum cut is defined capacity cf(u,v)=c(u,v)−f(u,v) and no flow. The
as a cut on graph G that has minimum cost. idea behind the Ford-Fulkerson method is that if
there exists a path from s to t in the residual
The min-cut/max-flow theorem states that for any network, then we can augment the current flow by
directed linear graph the maximum flow value this path. This path is called the augmenting path.
from s to t is equal to the cost of the minimum cut The augmentation of this path is equal to the
separating s and t . In other words, finding the smallest residual capacity along this path. Once no
minimum cut of a graph is equivalent to finding its augmenting path can be found, the current flow
maximum flow. must be the max-flow.
In order to illustrate the concept of flow in a graph, let
us interpret the directed graph as a network and the
edges as pipes connecting the nodes. Each pipe has a
certain capacity c{p,q} that corresponds to the weight
of the edge w{p,q}. Now we can push a
flow f(s,t)through the network leaving the source and
arriving at the sink. According to the min-cut/max-
flow theorem, fmax=|C|min.

Before stating the conditions that define a finite The Ford-Fulkerson method is called a method and
flow f in a network, let us denote all outgoing edges not an algorithm because it does not specify how
from node p and all incoming edges to node p by, one should go about finding the augmenting path.
O(p)={{p,q}∈E|q∈V} The Edmonds-Karp algorithm specifies that we use
I(p)={{q,p}∈E|q∈V} a breadth-first search to find a viable path in every
iteration.
respectively. The first condition is given by,
 Boykov-Kolmogorov[2]: - This method gives the
values that can be assigned as the edge weights.
The n-link edges must have weights carefully
computed in order to reflect inter-pixel similarities.
Concretely, we want the weight of an edge to be
big with the two pixels are similar, and small when
they are quite different. One idea is to let the
weight be boundary penalty, a function that maps
which is comparable to Kirchhoff's current law. two pixel intensities to a positive integer. Let Ip be
Assuming that outcoming flows are positive and the brightness, or intensity, of the pixel vertex p.
incoming flows are negative, the sum of all outgoing
For any edge (p,q)∈E, the boundary
and incoming flows must be zero for all nodes but the penalty B(Ip,Iq) is defined to be: -
source and the sink. The flow emerging from the
source, f, is equal to the flow arriving at the sink.
Secondly, capacities must be finite, i.e., c{p,q}<∞.
Finally, the flow within an edge cannot exceed its
capacity, f(p,q)≤c{p,q}.
Efficient Implementation of Max-Flow
 Regional Property: - This the property that helps
We have adopted the min-cut/max-flow algorithm in determining the edge weights for the source and
proposed in ,which is broadly used in the literature. It the sink of the graph,
is based on the augmenting path concept .The
algorithm works on a residual graph Gf, which is
initialized as G. In each iteration, a path along non
saturated edges from s to t is searched in Gf. The
smallest capacity along the path determines the
maximum flow Δf that can be pushed. The residual
capacities of the edges along the augmented path are
reduced by Δf, while the residual capacities of the
reverse edges are increased by the same amount. The
total flow from s to t is increased, f=f+Δf. The algorithm
terminates when there are no more s→t possible paths.

 Edmond-Karp: - The Edmonds-Karp algorithm


is an implementation of the Ford-Fulkerson
method. First, we define a residual
Algorithm: - Each pixel in the image is associated with the
Steps of Implementation: - nearest cluster center whose search area overlaps this
pixel. After all the pixels are associated with the nearest
 The object and the background seeds are marked. cluster center, a new center is computed as the average
 Each pixel of the image is considered as a vertex labxy vector of all the pixels belonging to the cluster. We
called as a pixel-vertex of the graph. then iteratively repeat the process of associating pixels
 Using the Boykov-Kolmogorov algorithm we with the nearest cluster center and recomputing the cluster
assign the edge-weights to the graph formed. center until convergence.
 We add two extra nodes called as the source(s) and
the sink(t) nodes.
 The values to these nodes is assigned based on the
regional properties.
 Now we have a complete graph with the source,
sink and all the other vertices.
 We apply the Edmond-Karp algorithm to find the
augmenting path for the assigned graph.
 We then segment the image along the augmenting
path.

III. OUR P ROPOSAL


 SLIC[3](Simple Linear Iterative
Clustering)]Segmentation: - This approach
generates superpixels by clustering pixels based on
their color similarity and proximity in the image
plane. This is done in the five-dimensional [labxy]
space, where [lab] is the pixel color vector in
CIELAB color space, which is widely considered
as perceptually uniform for small color distances,
and x,y is the pixel position. While the maximum  CIE Color Difference Formula[4]: - This formula
possible distance between two colors in the helps in calculation of the intensity, which in turn
CIELAB space (assuming sRGB input images) is helps in the formation of super pixels in the SLIC
limited, the spatial distance in the xy plane depends algorithm. The formula keeps in mind the two main
on the image size. It is not possible to simply use factors that differentiates the pixels from each
the Euclidean distance in this 5D space without other: - chroma and hue.
normalizing the spatial distances. In order to cluster
pixels in this 5D space, we therefore introduce a
new distance measure that considers superpixel
size. Using it, we enforce color similarity as well as
pixel proximity in this 5D space such that the
expected cluster sizes and their spatial extent are
approximately equal.
This algorithm takes as input a desired number of
approximately equally-sized superpixels K. For an
image with N pixels, the approximate size of each
superpixel is therefore N/K pixels. For roughly
equally sized superpixels there would be a
superpixel center at every grid interval
S = (N/K)1/2. At the onset of our algorithm, we
choose K superpixel cluster centers Ck = [lk; ak; bk;
xk; yk]T with k = [1;K] at regular grid intervals S.
Since the spatial extent of any superpixel is
approximately S2 (the approximate area of a
superpixel), we can safely assume that pixels that
are associated with this cluster center lie within a
2S x 2S area around the superpixel center on the
xy plane. This becomes the search area for the
pixels nearest to each cluster center.
Steps of Implementation: -
 Mark the background and the foreground for the
image.
 Calculate the intensity of the scribbles using the
CIE color difference formula.
 Call for the SLIC algorithm
 It will create clusters of similar
intensities
 Calculate the centroid of each cluster.
 Calculate the intensity of the centroid of each
cluster.
 Fill in the cluster with the above calculated
intensity resulting in the formation of the super
pixel.
 A histogram will be generated that will keep the
track of number of super pixels having the same
intensity (i.e. an intensity to number histogram).
 The edge weight of the graph will be equal to
100*e(-(difference in the heights of the histograms)/(distance
between the centroids ))
V Conclusion
 We will add two new nodes that will be source(s) As we can see from the above comparison the paper
and the sink(t). proposed methodology is better when there are only a
 The edge weights of the source and the sink will be single or few number of differentiating boundaries and
the 3D Euclidean distance from each of the vertex. the image is gray-scale but our proposed method holds
good for real life situations where there are a large
 We have our graph ready so we will apply the number of differentiating boundaries and the images are
Edmond-Karp algorithm to find the augmenting colored.
path.
 Cut the image along the augmented path. VI Future Work
Future work may include: -
IV RESULT COMPARISON (1) Improve the proposed method in case of overlapping
Paper Proposed Our Proposed Method of super-pixels which create discrepancies.
(2) Automatic marking of the foreground and
Method background scribbles.
(3)segmenting/identifying a particular thing from the
image. For eg;- identifying a particular person from an
image of a crowd, this can help in various
investigations.
(4)segmenting images that are clicked in the dark.

VII References
[1] Raquel Fandos, Leyna Sadamori and Abdelhak M. Zoubir ” HIGH
QUALITY SEGMENTATION OF SYNTHETIC APERTURE SONAR IMAGES USING
THE MIN-CUT/MAX-FLOW”
[2] Y. Boykov and V. Kolmogorov, “An experimental comparison of
min-cut/max- flow algorithms for energy minimization in vision,”
IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.
26, no. 9, pp. 1124 – 1137, 2004..
[3] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi,
Pascal Fua, and Sabine S•usstrunk.
[4] Gaurav Sharma, Wencheng Wu,Edul N. Dalal “
The CIEDE2000 Color-Difference Formula: Implementation
Notes, Supplementary Test Data, and Mathematical Observations”

IEEE conference templates contain guidance text for


composing and formatting conference papers. Please
ensure that all template text is removed from your
conference paper prior to submission to the
conference. Failure to remove template text from
your paper may result in your paper not being
published.

Вам также может понравиться