Section: New Results

Multi-scale spot segmentation with automatic selection of image scales

Participants : Bertha Mayela Toledo Acosta, Patrick Bouthemy.

Detecting spot-like objects of different sizes in images is required for many applications. A spot detection framework can be divided in three sub-steps : first, image preprocessing to smooth out noise; second, signal enhancement to highlight spots; third, spot detection by thresholding; the two first ones being often merged in a single operator. However, elements of interest do not all correspond to the same image scale, if the collection includes subgroups of different sizes or if perspective effects occur. Then, the need is not merely the selection of the optimal image scale, but of all the meaningful scales. We dealt with the problem of multi-scale spot detection while automatically selecting the meaningful scales. Our primary interest is to detect particles in microscopy images, but our method can be applied to other types of images as well. We defined an original criterion based on the a contrario approach and the LoG scale-space framework to automatically select the meaningful scales. We designed a coarse-to-fine multi-scale spot segmentation scheme involving a locally adaptive thresholding across scales, to come up with the final map of segmented spots. We carried out experimental results on simulated and real images of different types, and we demonstrated that our method outperforms other existing methods.

Reference:  paper accepted, ICASSP'2017.

Collaborator: Antoine Basset (CNES, Toulouse).