CV Workshop 3

MISCADA – computer vision

Workshop 3

Colour segmentation

Normalise the RGB values

• Please refer to https://en.wikipedia.org/wiki/Rg_chromaticity
• For a colour image of your choice
  • Scan all the pixels
  • For each pixel normalize its value
  • Use only the r and g normalised values
• Using Matplotlib, display the pixels in the [r,g] space

Implement the K-means

• This clustering algorithm is available, however I want you to develop your code for it, it is very simple
• To test the implementation use 200 random points in a 2D space with positive coordinates and no larger than 100

Steps

• Initialise K centres
• While centres change or a number of cycles have elapsed
  • Assign each pixel to the closest centre
  • Recalculate the centres using mean value

Normalised RGB space clustering

• Use the K-means algorithm to cluster a colour image
• Assume 4 colours
• Initialise the K centres at random in the [r,g] space
• Using Matplotlib, display the centres at initialization and at convergence
• Once the algorithm converged, colour image pixels using the 4 clusters
• Using Matplotlib, display original colour image and the clustered image

YUV space

• Using the K-means, cluster in the [u,v] space and image of your choice
• Follow the instructions from the previous slides and display the original colour image and the clustered one

Assessing your implementation

• To assess your implementation, you can use the available K-means algorithms
• It is readily available in the SciPy library
• SciPy K-means Documentation
• Do implement clustering in [r,g] and [u,v] using the SciPy K-means and then compare with your implementation