Noise sensitivity of watershed segmentation for different connectivity: experimental study

Low-budget image acquisition devices can produce noisy images, degrading segmentation quality. The sensitivity of two basic neighbourhood definitions to noise in watershed segmentation is compared. The two neighbourhood definitions are the 4- and 8-connected neighbourhoods. A similarity of segmentation to the noiseless original is observed, while increasing the energy of random Gaussian noise added to the image. It is concluded that the 8-connected neighbourhood is more robust.