We propose a new approach for image segmentation at different scales of observation, based on a multiscale image decomposition and on the active contour segmentation model. The proposed method consists of two steps. Firstly, a representation of a given image at multiple scales is derived, by means of a smoothing method which minimizes the weighted total variation norm of the image. This method allows the longtime preservation of edges and contrast with increasing scale, facilitating the detection of underlying structures. Secondly, image structures are extracted at each scale, using a level set formulation of active contours, minimizing the Mumford-Shah functional. Promising results of the proposed segmentation approach on natural images are reported.