Clustering-Based Segmentation
There is no broad hypothesis of image segmentation. Be that as it may, with the presentation of numerous new speculations and strategies for different controls, there
Original Image
Sobel Operator
Laplacian Operator
FIGURE 4.5 Examples of different operator images.
have been many image division techniques joined with some particular hypotheses and strategies. The purported class alludes to the assortment of comparative components. Clustering is as per certain prerequisites and laws of the arrangement of things all the while. The clustering method of the feature space is utilized to fragment the pixels in the image space with the compared highlight space focuses. As per their total in the feature space, the feature space is segmented, and afterward they are mapped back to the first image space to obtain the segmented result. There are two fundamental classifications of clustering techniques: hierarchical strategy and partition-based technique. Various hierarchical techniques depend on the idea of trees. In this, the root node of the tree represents the entire database, and the inner nodes represent different clusters. On the opposite side, partition-based techniques use enhancement strategies iteratively to minimize a work objective function. In the middle of these two techniques, there are different algorithms to discover clusters. There are essentially two sorts of clustering (Dehariya et al., 2010).
Hard Clustering
Hard clustering is a basic grouping strategy that partitions the image into a lot of groups with the end goal that one pixel can just have a place with just one cluster. As it were, it is also defined as that every pixel can have a place with precisely one cluster. These strategies use enrolment capacities having values either 1 or 0; for example, one either certain pixel can have a place with a specific cluster or not. A case of a hard clustering-based strategy is one k-mean clustering-based method known as hard clustering mean. In this method, above all, the focuses are figured, and then every pixel is relegated to the closest focus. It stresses on augmenting the intra-clus-ter similitude and furthermore limiting the between cluster balance.
Soft Clustering
Soft clustering is a progressively specific type of clustering because precise segmentation in real life is not possible due to the occurrence of noise. Thus, the soft clustering technique is most useful for image segmentation in which segmentation is not strict. An example of such a technique is fuzzy c-means clustering. In this technique, the pixels are divided into clusters based on partial membership, that is, a pixel can belong to more than one cluster and is described by membership values of this degree. This technique is more flexible than other techniques (Yamal and Gupta, 2013).
Some types of clustering techniques are most commonly used in image segmentation.
K-Means Clustering Technique
K-means is one of the most commonly used clustering algorithms. The basic idea of K-means is to collect samples in different groups according to the distance. The closer the two points are to achieving compact and independent clusters as close as the intended targets (Suleiman and Isa. 2010). The implementation process of K-means is expressed as follows:
- 1. randomly select K initial clustering centers;
- 2. calculate the distance of each cluster center from each sample and return each sample to the nearest clustering center;
- 3. for each cluster, through all samples as clusters of new clustering centers;
- 4. repeat steps (2) to (3) until the cluster center no longer reaches or reaches the specified number of changes (Kumar and Singh, 2013).
The advantage of the K-means clustering algorithm is that the algorithm is fast and simple and is highly efficient and scalable to other data sets. Its time complexity is close to linear and suitable for large-scale data set mining. The disadvantage of K-means is that its clustering number K has no clear selection criteria and is difficult to estimate (Chuang et al., 2006). Secondly, it can be seen from the K-means algorithm framework that each iteration of the algorithm traverses all samples, so the algorithm time is very expensive. Finally, the K-means algorithm is a distance-based segmentation method (Celebi et al., 2013). This only applies to the data set that is convex and not suitable for adding nonwave convex groups (Figures 4.6 and 4.7).
FIGURE 4.6 Examples of different K-means images (k = 2).
Original Image
Segmentation 1
Segmentation 2
Segmentation 3
FIGURE 4.7 Examples of different K-means images (k = 3).
Fuzzy C-Means Clustering Technique
The fuzzy c-means algorithm is an unheard fuzzy clustering algorithm. The traditional clustering algorithm finds a “hard partition” of a given dataset based on certain criteria that evaluate the goodness of the partition. By “hard partition.” we mean that each datum belongs to exactly one cluster of partitions, whereas the soft clustering algorithm finds the “soft partition” of a given dataset. In “soft partitions,” the datum may be partially related to several groups. A soft partition is not necessarily a fuzzy partition because the input space can be larger than the dataset. However, most soft clustering algorithms generate a soft partition that also creates a fuzzy partition. A type of soft clustering of particular interest is one that ensures the degree of membership of point x for one, that is, in all groups.
(4.1)
A soft partition that satisfies this additional condition is called a constrained soft partition. The fuzzy c-means algorithm, which is the best-known fuzzy clustering algorithm, produces constrained soft partition. In order to produce constrained soft partition, the objective function J, of hard c-means has been extended in two ways:
- 1. The fuzzy membership degree in cluster has been incorporated in the formula.
- 2. An additional parameter m has been introduced as a weight exponent in fuzzy membership. The extended objective function, denoted by J„„ is:
(4.2)
where P is fuzzy partition of dataset X formed by C_{t}.C_{2},....., C_{k} and k is the number
of clusters. The parameter m is weight that determines the degree to which partial members of cluster affect the clustering result. Like hard c-means, fuzzy c-means also tries to find good partition by searching for prototype v, that minimizes the objective function J_{m}. Unlike hard c-means, however, the fuzzy c-means algorithm also needs to search for membership function n_{Ci} that minimizes J_{m}. A constrained fuzzy partition {C,,C_{2},.....,CJ can be the local minimum of the objective function
J_{m} only if the following conditions are satisfied:
Mc_{(}(x) =
(4.3)
Where 1 < i < k. x e X
(4.4)
where 1 < i < k
Few important points regarding the FCM algorithm: It guarantees converge for m > 1. It finds local minimum of the objective function J,„. The result of applying FCM to a given dataset depends not only upon the choice of parameter m and c but also on the choice of the initial prototype.
The fuzzy c-means clustering algorithm is applied on a MRI cerebral image. The segmentation output presented in the figure below Figure 4.8(a) corresponds to a human cerebral cut; it is the original input image of the program which is split into (w x n) elementary images as in Figure 4.8(b). Figures 4.8(c)-(e) are the segmented output images where each of them corresponds to a specific class: the gray matter, the cerebrospinal fluid, and the white matter.
a) Input Big Data Image
b) Elementary Images
- c) Gray Matter
- d) Cerebrospinal Fluid
- e) White Matter
FIGURE 4.8 Segmentation results by the elaborated distributed program.
Segmentation Based on Weakly Supervised Learning in CNN
Learning in CNN
In recent years, image classification, detection, segmentation, high-resolution image formation, and many other fields have found breakthrough results in deep learning (Zhang et al., 2013). In the aspect of image segmentation, an algorithm has been proposed that is more effective in this area, which is the weakest and semi-fabricated learning of DCNN for arithmetic image segmentation. Google's George Papandreou and UCLA’s Liang-Chieh Chen studied the use of bounding boxes and image-level labels as markup training data based on DeepLab and the expected maximization algorithm to estimate predicted pixel squares and CNN parameters (Em) used. The DeepLab method is divided into two phases, the first step is still using FCN to obtain a coarse score map and project the size of the original image, and then the second stage is the CRF fully connected to the FCN Borrowing to obtain the details of segmentation refinement (Chen et al., 2017) (Figure 4.9).
FIGURE 4.9 DeepLab model training using image-level labels.
FIGURE 4.10 DeepLab model training from bounding boxes.
For image-level tagged data, we can observe the image’s pixel value x and the image-level mark z. but do not know the label y for each pixel, so y is considered as a hidden variable. Use the following probability graph model:
/ I I M A
P(x,y,z;0) = P(x)l £ £ P(y_{m}x-,6) jp(zly) (4.5)
Use the EM algorithm for estimation and y. The E step is fixed with the value of Y expected, and the M step is fixed to calculate Y using SGD (Papandreou et al., 2015) (Figure 4.10).
The training image that returns the bounding box mark uses the CRF to automatically segment the training image and then perform full supervision based on segmentation. Experiments show that using the image level of the bus sign to achieve the segmentation effect is poor, but better results (Fu & Mui, use of bound box training data, 1981) can be found.
Comparative Study of Image Segmentation Techniques
These methods are comparatively studied using some standard parameters such as spatial information, field continuity, speed, computation complexity, automaticity, noise resistance, multiple object detection, and accuracy. Table 4.1 presents an analysis of all methods.
co
O
TABLE 4.1
Comparison of Image Segmentation Methods
Parameters |
Spatial Information |
Region Continuity |
Speed |
Computation Complexity |
Automaticity |
Noise Resistance |
Multiple Object Detection |
Accuracy |
Threshold |
Ignored |
Reasonable |
Fast |
Less |
Semiauto |
Less |
Poor |
Moderate |
Region-based |
Considered |
Good |
Slow |
Rapid |
Semiauto |
Less |
Fair |
Fine |
Cluster |
Considered |
Reasonable |
Fast |
Rapid |
Automatic |
Moderate |
Fair |
Moderate |
Fuzzy C-mean |
Considered |
Good |
Moderate |
Moderate |
Automatic |
Moderate |
Fair |
Moderate |
Advanced Digital Image Processing and Its Applications in Big Data