In this paper, a novel graph theoretic image segmentation technique is proposed, which utilizes forest concept for clustering. The core idea is to obtain a forest from the image followed by construction of average value super pixels. Thereafter, a merging criterion is proposed to merge these super pixels into two big classes thereby binarizing and thresholding the image separating background from foreground. Extensive experimentation and comparative analysis are finally performed on a diverse set of images to validate the technique and have noted the significant improvements. © Springer Nature Singapore Pte Ltd 2020.