In this paper, a 2D image, stored as a 2D array of pixels, is represented as a function f: Z2 yields Z such that f(x) is the gray scale value of the image function f at the point x (epsilon) Z2. The set of points (x,f(x)) where x (epsilon) Z2 can be considered as a topographical surface S. We formally define `blob-like' features on this surface S as regions associated with extrema of f. We associate a blob-like feature to each extremum of the surface S. The blob-like regions have spatial as well as gray level extent. These features are the connected components of the surface S. Features associated with minima correspond to darker objects on brighter background whereas those associated with maxima correspond to brighter objects on darker background. It is well known, that not all image features can be detected at any one scale. We therefore propose a multi- scale scenario wherein the complete information about all the image features is contained in the gaussian scale space representation of that image. We detect the formalized `blob-like' features at each scale of the scale space and translate the problem of representation of these features across the scales to a tree representation G. The nodes of G correspond to the blob like features and the relative positioning of the gray level blobs in the tree is decided by the topographical surface S. These blobs are grouped into a higher order image structure, called the (sigma) -blob which is a family of blobs over an interval of scale. The detection of the prominent image features is translated to the problem of the traversal of the (sigma) -blob tree which is the representation of the (sigma) -blobs as nodes and the relative positioning in this tree representation is decided by the scale at which these (sigma) -blobs would be detected. Additional properties of the blob-like features are established both at single scale and across the scales. The detection algorithm proposed in the work is presented with test runs on both synthetic and real images.