Classification of Image Segmentation Techniques

5 min readMay 28, 2021

Image segmentation is the process of dividing an image into set of pixels to make the image less complex. Pixels within a set have one or more attributes (texture, intensity, color) in common.

Thresholding methods-

1.1 Global Thresholding-

This method is used when the object are easily differentiated from each other, so we can use a single value as threshold for the entire image.

Threshold value should not be too high or too low, it must be optimal.

For binary image if pixel value less than threshold value it converts that pixel value to black else it converts into white.

1.2 Local Thresholding-

Local thresholding can be defined as:

Where, g(x, y) - binary image

I(x, y) - intensity of the each pixel

T(x, y)-threshold value

This method decides multiple threshold values for every pixel in the image on the basis of attributes (range, variance or surface-fitting parameters) of adjacent pixels.

As we can set multiple threshold values in local thresholding, this method works well on high grayscale contrast images where global thresholding method will not work effectively.

2.Clustering-based Segmentation-

In clustering-based segmentation the pixels in the image are divided into groups, where some property of the pixels in each group is similar.

Clustering-based segmentation use K-means algorithm. This algorithm for Image Segmentation helps to enhance high performance and efficiency. The user has to specify the number of cluster. Selection is based on number of clusters determined using datasets from images by using frame size and the absolute value between the means of clusters.

3.Edge-based Segmentation-

Edges are defined as sudden change of intensity levels in a digital image. This technique is based on discontinuity in an image. Edge detection is used to detect the boundaries or to find size or location of an object in an image.

Edge detection techniques can be further classify as-

1.Gradient Based: Calculates first-order derivative

2.Guassian Based: Calculate second-order derivative

1.1 Sobel Edge Detection-

Sobel operator calculates the gradient approximation of image intensity function to detect edges. The following kernels are used for convolution with the input images.

This method detects smooth edges easily and is simple and time efficient. It does not accurately detects thick and rough edges and also doesn’t preserve diagonal direction points. It has high noise sensitivity.

1.2 Prewitt edge detection-

The orientation and magnitude of an image is detected by this method. It detects the vertical and horizontal edges of an image. It uses following kernels for convolution with the input images.

Prewitt is quite similar to the sobel edge detection technique, but it is a bit easy to implement than sobel. This operator can sometimes generate noisy results.

1.3 Robert edge detection-

The sum of squares of the differences between diagonal pixels are calculated through discrete differentiation. After which the gradient approximation is decided. The following 2x2 kernels for convolution with the input images.

It detects the edges and orientation easily while preserving the diagonal direction points. It has high noise sensitivity.

2.1 Canny edge detection-

This is optimal edge detection technique as it selects image features without impacting the feature. Noise doesn’t has any effect on this technique. Edges are detected on a basis of low error rate, accurately localized edge points, single edge response.

2.2 Laplacian of Gaussian-

This method is also known as Marr-Hildreth operator. It is a Gaussian based method where Gaussian operator reduces noise and the Laplacian operator detect sharp edges. It is used when there are sudden grey level transitions. An edge is detected when the second-order derivative crosses zero, this is also called as zero-crossing method.

Gaussian function-