The equations that govern geophysical fluids (namely atmosphere, ocean and rivers) are well known but their use for prediction requires the knowledge of the initial condition. In many practical cases, this initial condition is poorly known and the use of an imprecise initial guess is not sufficient to perform accurate forecasts because of the high sensitivity of these systems to small perturbations. As every situation is unique, the only additional information that can help to retrieve the initial condition are observations and statistics. The set of methods that combine these sources of heterogeneous information to construct such an initial condition are referred to as data assimilation. More and more images and sequences of images, of increasing resolution, are produced for scientific or technical studies. This is particularly true in the case of geophysical fluids that are permanently observed by remote sensors. However, the structured information contained in images or image sequences is not assimilated as regular observations: images are still (under-)utilized to produce qualitative analysis by experts. This paper deals with the quantitative assimilation of information provided in an image form into a numerical model of a dynamical system. We describe several possibilities for such assimilation and identify associated difficulties. Results from our ongoing research are used to illustrate the methods. The assimilation of image is a very general framework that can be transposed in several scientific domains.