Week 2 of summer internship at USM
· I In order to perform the classification of greyscale images, I found that the Fuzzy C Means (FCM) algorithm which assumes a membership degree for each pixel for every cluster, a good starting point.
· However, the FCM suffers from the drawback that an estimate of the number of clusters needs to be provided for the algorithm to perform classification
· To sort out this problem, it was proposed to initialize the FCM with the resultant number of clusters obtained by an initial segmentation performed by the Mean Shift Algorithm (MS)
· It was noted that MS suffers from over segmentation and the quality of classification largely depends on the assumed neighbourhood distance. It would be appreciable if one could automatically decide the optimal neighbourhood distance given an image.
· The FCM was initialised with the number of clusters given by MS and the results were obtained.
· It was noted that the FCM does not incorporate any spatial information, such as the relationship of a pixel with its neighbourhood. The degree membership function could be modified to incorporate this information. It is expected that the results should improve.
The proposed method (without the incorporation of the spatial information) was run on two images –a synthetic image (Aluminium Grains: 149 x 100) and a natural image (Lena 100 x 100.) Different neighbourhood distances (NDs) were selected for performing MS. Following are the results for a particular case for each of the images:
Alum Grains -Original
Mean Shift with ND=50 giving 8 clusters
FCM initialized with 8 clusters
Lena -Original
Mean Shift initialized with ND=40 giving 4 clusters
FCM Initialized with 4 clusters