Object extraction in an image is a challenging problem since segmenting an object needs a combination of various visual cues such as texture information, distribution of the pixel intensities, object boundaries and dissimilarity from other objects/regions. Furthermore, the visual scenery may bear more than one object, or partially occluded objects. Unfortunately, image segmentation that can be defined as extracting meaningful regions from an image is an ill-‐posed problem since one has to first define and fix a priori what a “meaningful region” is. As a result, current automatic object segmentation algorithms suffer from various problems that make them unreliable and deficient for the purpose of (salient) object extraction. Learn-‐to-‐segment (L2S) proposes a novel operational neural network (ONN) in order to improve the traditional deep learning neural networks, in which “convolution” is the only operator used. We believe that extending the set of operations, both linear and non-‐linear, can lead to improving the generalization capability of the neural network. However ONNs, like their predecessor CNNs, will be classifiers that are proven to localize the object(s) in an image and can (learn to) extract the features for the purpose of image (object) classification. Therefore, we shall then modify the ONN topology, its training and outcome so as to serve to the L2S purpose to yield Operational (Object) Segmentation Networks (OSNs) where the output of a OSN will be the saliency map of the image from which salient object(s) can be segmented.