Sparse coding of natural images using a prior on edge co-occurences

Architecture of the model
Figure 1: Edge co-occurrences (A) An example image with the list of extracted edges overlaid. Each edge is represented by a red line segment which represents its position (center of segment), orientation, and scale (length of segment). . (B) The relationship between a reference edge "A" and another edge "B" can be quantified in terms of the difference between their orientations, ratio of scale, distance between their centers, and difference of azimuth. This is used to compute the chevron map in Figure 2. Go back to manuscript page.

Sparse Coding Of Natural Images Using A Prior On Edge Co-Occurrences
Oriented edges in images commonly occur in co-linear and co-circular

arrangements, obeying the good continuation law of Gestalt psychology. The human visual system appears to exploit this property of images, with contour detection, line completion, and grouping performance well predicted by such an association field between edge elements \citep{Field93,Geisler01}. In this paper, we show that an association field of this type can be used to enhance the sparse representation of natural images. First, we define the SparseLets framework as an efficient representation of images based on a discrete wavelet transform. Second, we extract second-order information about edge co-occurrences from a set of images of natural scenes. Finally, we incorporate this prior information into our framework and show that it allows for the extraction of features relevant to natural scenes, like a round shape. This novel approach points the way to practical computer vision algorithms with human-like performance.


slides, slides with notes


For more information on Matching Pursuit on natural images, follow The hitchhiker guide to Matching Pursuit.


All material (c) L. Perrinet. Please check the copyright notice.

This work was supported by European Union project Number FP7-269921, "BrainScales".
BrainScaleS logoFET logoFP7 logoEU logo

TagYear15 TagBrainScales TagPublicationsArticles TagTalks

welcome: please sign in