A study of the effect of medial-axis representation on robust pattern recognition

Abstract

There are experimental findings suggesting that the human visual system acquires the medial- axis representation (skeleton) of an object (Kovacs & Jules, 1993, 1994). If it is true, there is the possibility that the medial-axis representation has some computational advantage. In order to examine the possibility that the merit resides in the robustness of pattern recognition, this study compared the cases when the outer edge contours of various objects were inputted to the CNN (deep convolutional neural network), and the case when medial-axes extracted from the same object set were inputted. For each case, we investigated the relationship between the degree of various deformations such as skew-shearing given to the pattern and the magnitude of subsequent change on the recognition ability. The results indicated that the medial-axis representations by themselves did not bring so superior robustness, suggesting that they might be used with other contour cues as the supplemental information, or contributing other computational aspects besides robustness.