image segmentation and defect detection techniques using homogeneity
TRANSCRIPT
IMAGE SEGMENTATION AND DEFECT DETECTION TECHNIQUES USING HOMOGENEITY
黃偉鑫
Outline Abstract Local homogeneity analysis and H-image Homogeneity for image segmentation Homogeneity for image defect detection Conclusion & Experiment result
Abstract survey on latest image segmentation and defect
detection techniques using local homogeneity analysis
image segmentation suffer from the problem local homogeneity based approaches for image
segmentation and image defect detectionregion merging Gabor filteringdiscrete cosine transform color features wavelet transform watershed algorithmFFD method Information Fusion
Local homogeneity analysis Criterion definition
Let P be the square window of width 2 N + 1
f be the sum of all the vectors defined in P
Local homogeneity analysis H value
Local homogeneity analysis how to determine the size of the local window
large windows are useful for detecting texture boundaries
small windows are useful in localizing the intensity edges
H-image gray-scale image
image segmentationHomogeneity With Region Merging
avoid the over segmentation faced the problem of selecting appropriate local
scale for determining the most appropriate local window sizeSeeded Region
Determination
Region Growing
Region Merging
image segmentationHomogeneity With Region Merging
Seeded Region Determination Calculate the average and the standard deviation of the H-
values in the H-imageLet them be and For each pixel p in H-image, compute the average and the standard deviation of pixels in the local window set a threshold Tp in p
If the H-value of p is less than T, p is considered as a candidate seeded region point (CSRP)
image segmentationHomogeneity With Region Merging
Region Growing average the H values of all the points that do not belong to any
seeded region connect pixels below the average assigned to their neighboring seeded regions
Region Merging For color images, the regions are merged based on their color
similarity For gray-scale images, although the merging can be based on
texture features The merging process continues until a maximum threshold for
the distance is reached
image segmentationHomogeneity with Information Fusion
a novel approach to natural scene segmentation
uses both color and texture features
image segmentationHomogeneity with Information Fusion
Feature Extraction suitable for inputting into the SOM for training and
subsequent classification each pixel is represented as a seven dimensional
vector {r, g, b, e5l5, e5s5, r5r5, l5s5} R, G, B values and Laws’ texture energy measures
image segmentationHomogeneity with Information Fusion
Selection of Training Samples non-homogeneity measureIn order to be used in color image, for the pixel location (i , j)
The average non-homogeneity value is calculated for the entire image
Local average non-homogeneity value for each block
image segmentationHomogeneity with Information Fusion
HSOM Training and Testing hierarchical two-stage self-organizing mapping
Region Merging output of the HSOM is often an over-segmented
image converted into the CIE color space
image segmentationHomogeneity With Color Features
This approach for color image segmentation which uses local homogeneity analysis and color features
modified the region merging approach
image segmentationHomogeneity With Color Features
Self-Constructing Fuzzy Clusteringeach pixel is associated with a color signal (x1, x2, x3) where x1, x2, and x3 denote the R, G, and B valuesCluster Cj is represented by (G1j,G2j,G3j) whereG1j , G2j , andG3j are Gaussian membership functions of R, G, and BEach Gij ,1 ≤ i ≤ 3, has the center mij and standard deviation σij
we create a new cluster Ck, k = J + 1, withmik = xi, σik = σ0
where 0 ≤ ρ1 ≤ 1 is a predefined thresholdd(x1, x2, x3;Cj) ≥ ρ1
Otherwise, let cluster Ct be the cluster with the largest value of similarity d
image segmentationHomogeneity With Color Features
Color-Based and H-based Region Growing A color-based region is formed by connecting a
set of pixels with the same color according to the spatial relationship of eight-connectivity
we find out those pixels with high local homogeneity by H-value calculation, group them into several seeded regions, and obtain a seeded region image
image segmentationHomogeneity With Color Features
Region-Based Region Growing we grow the regions based on the color-based
regions which are labeled as non-seeded calculate the Eucliden distance of colors between
each non-seeded color-based region and each seeded region we combine the non-seeded region with the
lowest distance into the corresponding seeded region and update the color and size of the seeded region
image segmentationHomogeneity With Color Features
Merging merge regions by color similarities
the pair of neighboring regions with the smallest dissimilarity are merged together to form a new region if the dissimilarity is smaller than a threshold
Combine the smallest region until all sizes of regions are bigger than a
threshold
image segmentationHomogeneity With Watershed
Transformation
watershed algorithm is that it produces many segments which make it difficult to merge the segments correctly
region merging cost plays an important role in the region merging process
Homogenei
ty
Watershed
segmentation
Region Mergin
g
image segmentationHomogeneity With Watershed
Transformation Region MergingRegion adjacency graph (RAG), G = (R,E,C) R denotes the sets of nodes(regions)E denotes the sets of edgesC denotes the sets of merging cost ofthe arbitrary pair of adjacent regions
For a pair of nodes rj and ri
image segmentationHomogeneity With Watershed
TransformationThe color homogeneity
The circle-like degree homogeneity
The rectangle-like degree homogeneity
l is the perimeter of the region
hcom represents the cluster degree of the pixels in the region
b is the perimeter of region bounding boxhsth represents the smoothness degree of the region boundary
The weights wcolor , wshape , wcom and wsth are the inputparameters
c the standard deviation within region of channel cThis value indicates the similar degree of the two adjacent regions
image segmentationHomogeneity With Watershed
Transformation
image segmentationHomogeneity With Gabor Filtering
Gabor filtering is a wavelet based method to extract texture features of a multispectral image
The approach worked very well for many multispectral images
GABOR FILTERING
FCM algorithm
Local homogene
ity analysis
image segmentationHomogeneity With Gabor Filtering
The Fuzzy C-means (FCM) clustering algorithm1.Suppose x=(x1, x2,…,xj....,xN) the pixels of the image to be partitioned. We choose the number of cluster, the fuzzification parameter m and the stopping condition ξ2.Initialize u using random value in range [0,1]3.Set the loop counter b =04.Calculate the cluster centre vi :
5.Update uij as a function of distance to clusters:
6.If Max{u(b)-u(b+1)}<ξ then stop, otherwise, set b=b+1and go to step 4
image segmentationHomogeneity With Gabor Filtering
Image defect detectionHomogeneity With FFD Model
used homogeneity and free form deformation model (FFD) for the attrition detection in banknotes
defect localization
Image defect detection Homogeneity with Discrete Cosine
TransformThese five energy measures are EH, EV, ED, EM and ES which are horizontal, vertical, diagonal, average and the standard deviation of energy value of a DCT block respectively
Image defect detection Homogeneity with Wavelet Transform
a new defect detection algorithm based on local homogeneity and hotteling model to localize defects in various textures images
Conclusion & Experiment result
optimal image segmentation for each images select local window size still rely experiment
result noise can also be treated as defect
Q&A