The standard FCM method did not yield satisfactory results with noisy images. It was tested with the Parrots image (300x300) corrupted by salt and pepper noise. The FCM method initialised by Mean Shift left the image misclassified and the noise was left uncorrected.
Hence, Spatial FCM was implemented which incorporates information about the 8-neighborhood for each pixel. It basically obtains the correlation of the membership function of a pixel with the probability that its neighbours belong to the same cluster as itself. The method yielded better results when initialised by Mean Shift. The results are shown below:
Further, it was noted that the choice of Mean shift bandwidth greatly affects the initialization procedure for FCM. I contacted Adel, and he described the problem to be still under research.
Parrots (300x300) corrupted by salt and paper noise.
Image, classified by FCM, initialised by mean shift (bandwidth=70, giving 8 clusters)
Image, classified by FCM-Spatial, initialised by mean shift (bandwidth=70, giving 8 clusters)
No comments:
Post a Comment