The overall operating principle of the SIFT-SVM is shown in Fig. 12.1. In the beginning level, preprocessing takes place to enhance the input image quality. Then, the image undergoes segmentation process by the use of the К-means clustering technique. Followed by, SIFT based feature extraction takes place, which extracts the needed features from the segmented image. At last, the SVM model is applied to classify the set of images into appropriate classes.

Bilateral Filtering–Based Preprocessing

Bilateral filtering (BF) is accepted as a preprocessing play to extract the noise that occurs in the dermoscopic images. Usually, the occurrence of noise leads to

Block diagram of SIFT-SVM

FIGURE 12.1 Block diagram of SIFT-SVM.

inefficient classifiers of the images. It can also be supposed for discriminating the noise that occurs in the initial image to the classifier. The possibility of BF is relied on a particular weight of connecting pixels to remove the noise. The simplest method for signifying a BF has the distance-based domain filter part (p, p' and a gray-value depending on range filter part r [(fp), (Jp'):

Where p and p' indicates the place of the intermittent and neighboring pixels, and N(p) is a normalized factor.

Regarding the local mean of the neighboring pixels, the range filter part implements the value-based module to remove the noise around the boundaries. The field and range filter areas, usually a Gaussian function, is utilized and depends on the Euclidean pixel distance as denoted as

where a,/ indicates the width parameter of the filter kernel and a, is the noise standard deviation of the regarded regenerated value.

Image Segmentation

It can be a clustering method that classifies or clusters the group of objects in terms of the attributes or features to the count of К sets, where К indicates the positive integer. The clustering method is transferred by minimizing the distance between the information and particular cluster centroid. A distance that is utilized here is Lr distance and is defined as (d(u, v) = "S' (u, - v )2)- The clustering play aims to


cluster the information in that matching objects in a cluster and objects of dissimilar clusters are not equal. The procedure involved in the К-means method is listed here.

  • 1. Setting the value of К that is regarded as centroids, i.e., the virtual points that have been created randomly.
  • 2. All points in the data set are shared to its nearest centroid.
  • 3. The position of the centroid is upgraded with the assignment of the data points to the cluster. Otherwise, the centroid is shifted to the center of its shared points.

The steps 2-3 are repeated until no centroids are moved to the subsequent round. In addition, the process gets terminated when the shift goes beyond a threshold value. The К-means method is then defining К groups of data that minimizes the subsequent objective function

where there exists К clusters Sp, p = 1, 2.....K, and c, is the centroid or mean

point of each point, uc/ e Sp. It is very useful in computer vision areas to segment the images. Each pixel that occurs in the image is connected to color as determined in RGB. The input image that requires to be segmented is defined utilizing a group of points from 3D data space. For a grayscale image, the method is equal, except that the illustration of images occurs as the gathered of points in a ID space.

Feature Extraction

This section looks at a SIFT-based feature extraction concept for extracting and explaining the feature points that require to be robust to scaling, orientation, and alteration in illumination. The group of four methods contained in the SIFT technique are:

  • 1. Detect scale-space extrema
  • 2. Localize feature points
  • 3. Assignment of orientation
  • 4. Feature point descriptor

Detect Scale-Space Extrema

Primarily, an exploring method occurs over the scale with the utilization of a Difference of Gaussian (DoG) application to search important interest points that are invariant for scaling as well as rotation. Now, scaling space is expressed as the function L(u, v, a) that is generated from the convolution of a variable-scale Gaussian G (u, v, o) containing an input image IMG(u,v):

For efficient detection of stable key-point locations from scale space with a function of scale-space extrema in the histogram of oriented gradients (HOG) function through the image, D (и, v, o) is defined from the difference of two closer scales that undergo division with constant factor к:

Localized Feature Points

A position and scaling of all interested points is calculated based on stability values that have been estimated with feature points. It is used to improve the contrast level of the images which causes the unclear image boundaries.

Assignment of Orientation

A multiple orientation has been chosen for every feature point location that depends upon the local image of gradient directions. For each image sample of L(u, v), the gradient magnitude m(u, v) and orientation Q(u,v) are predefined with the utilization of different pixel values

SIFT descriptors in a skin lesion image

FIGURE 12.2 SIFT descriptors in a skin lesion image.

Feature Point Descriptor

A feature descriptor is an approach which considers an image and offers feature vectors. The feature descriptors encode interesting details into a sequence of numbers and plays as an arrangement of numerical “fingerprint” that can be used to differentiate one feature from another. Fig. 12.2 depicts the SIFT descriptors in a skin lesion image.

Image Classification

The SVM is a type of ML method based upon the models of statistics. It can be a supervised learning method executed to identify designs in various fields. After the set of discriminate features has been selected with the earlier methods, the classification is executed to classify the various kinds of skin lesions. SVM is a supervised learning model employed for data testing, model reorganization, classifier, and regression testing. The SVM trained method generates the model that shares a novel instance through related class. In SVM, the linear function is employed so that the instances of the distinct class are separated by a clear gap. Given a train set of instance pairs (lp, mr), p = 1, -,/t, where /,, e R" and m e {1, -1}", the SVM requires the result of the following optimization problem:

where training vectors /,, are mapped into the highest dimensional space through the function ф.

Algorithm: SVM model

Input: T raining Image Features, Testing Image Features.

Assign: 6=resemblance among every sample that exists in the attribute.

Output: Classified Image.


Step 1: Give data set to the model

Step 2: Features and attributes undergo classification concerning labeled class

Step 3: Candidate Support Value evaluation Step 4: While (Instance_value! =NULL)

Step5: Reiterate the process for every instance Step 6: If (Support_value=8)

Step 7: Compute Total Error Value

Step 8: End If

Step 9: If (lnstance<0)

Step 10: Determine the Decision value = Support Value/Total Error

Step 11: Reiterate the processes till empty

Step 12: End If


An SVM determines a linear separating hyperplane with maximal margin in this maximum dimensional space. C > 0 is the punishment attribute of the fault term. Likewise, A"(lp,/(;) = ф(/;,)2ф(/,;) is called the kernel function. It uses the usually implemented Radial Basis Function (RBF) determined as

In the testing procedure, the input test image is segmented and the hair extraction method occurs. Next, features are removed from the image, and the following classifier method is carried out where the input images are classification into various classes.

< Prev   CONTENTS   Source   Next >