An information theoretic approach to reflectional symmetry detection
Agata Migalska, John P. Lewis · 2015
Symmetry is an omnipresent transformation in both nature and man-made objects. It is remarkable how human beings are capable of detecting and recognizing symmetries in the surrounding world without hesitation and without much apparent mental effort. On the other hand, teaching a machine to perform the same task has been challenging, resulting in a variety of approaches and algorithms. In this paper we appeal to information theory to obtain a novel and general principle for symmetry detection. Folding an image in half along a line that coincides with an axis of reflectional symmetry preserves the statistics of the image, whereas folding along any other line alters these statistics to become more Gaussian. Symmetric transforms can thus be detected as those that have the largest negentropy. Experimental evaluation shows that our algorithm properly detects the symmetry axes within synthetic and natural images and is applicable for reflectional symmetry of an arbitrary order.