Section: New Results

Small Particle Detection

Participants : Nicolas Cedilnik, Xavier Descombes, Eric Debreuve, Florence Besse.

This work is made in collaboration with Fabienne De Graeve (iBV).

One task of the RNAGRIMP project is to detect RNA granules from fluorescent images. These granules have their size close to the image resolution, they typically represent very few pixels. At this scale, shape parametric models are only crude approximations of the object geometry and not adapted for a detection task. To overcome this difficulty we have defined a shape dictionary consisting of all the shapes included in a five by five tile and satisfying some properties of regularity and convexity. Then we are mimicking the marked point process framework by defining an energy function on the finite sets of shapes as the sum of a data term, applied on each object, and a non overlapping constraint between neighboring objects. The solution minimizing the energy is approximated by a greedy algorithm. We have compared different data terms and shown better performances than the traditional threshold approaches and the wavelet based approach as provided by the software Icy.