MAPS: Multiscale Attention-based Pre-Segmentation of Color Images

Image segmentation is an essential preprocessing step towards scene understanding in computer vision. It consists in partitioning the image into connected regions which fulfill certain homogeneity criteria. Numerous segmentation techniques have been reported in the literature. Most of these techniques aim, however, at segmenting the entire image regardless of the relevance of each region. Furthermore the segmentation methods often use the same homogeneity criteria for all image regions, thus neglecting the feature-related specifcity of image segments. This paper reports a novel Multiscale Attention based Pre-Segmentation method (MAPS), which addresses the segmentation issues mentioned above. Inspired from psychophysical findings, our method is built around the multifeature, multiscale, saliency based model of visual attention. From the saliency map, provided by the attention algorithm, MAPS first derives the spatial locations that will be considered further in the segmentation process. Then, the method explores the model scale and feature space and extracts, for each salient location, the optimal scale an feature map required for presegmentation. This innovative presegmentation but yet uncomplete procedure must be followed by some refined segmentation that operates in the optimal feature map at full resolution.