Summary
We address here the problem of scale and rotation invariant object recognition, making use of a correspondence-based mechanism, in which the identity of an object represented by sensory signals is determined by matching it to a representation stored in memory. The sensory representation is in general affected by various transformations, notably scale and rotation, thus giving rise to the fundamental problem of invariant object recognition. We focus here on a neurally plausible mechanism that deals simultaneously with identification of the object and detection of the transformation, both types of information being important for visual processing. Our mechanism is based on macrocolumnar units. These evaluate identity- and transformation-specific feature similarities, performing competitive selection of the alternatives of their own subtask, and cooperate to make a coherent global decision for the identity, scale and rotation of the object.
See the full content of this document
Extract
A Visual Object Recognition System Invariant to Scale and Rotation
(ProQuest: ... denotes formulae omitted.)
1. IntroductionIn this work we address the problem of visual object recognition invariant to change in scale and orientation. This problem is far from being resolved for varying stimuli of natural complexity [1]. Following Rosenblatt [2, 3], invariance is often assumed to be achieved feature-by-feature, by connecting all cells of an input domain that correspond to transformed versions of a given feature type to one master neuron to represent presence or absence of that feature type. Recognition is then based on the set of feature types that occur in the image. This approach is called feature-based recognition. When changes in scale and orientation are among the allowed transformation, and when neurobiologically realistic feature types such as Gabors [4, 5, 6] are used, this approach runs into the difficulty that the main distinction between different features -scale and orientation- is completely lost, so that recognition breaks down. There are feature types that are immune to this difficulty [7], but these features are hardly able to express the texture of objects, such is necessary, for instance, for face recognition.An alternative approach, which we follow here, is to recognize the transformation between input image and stored representation and to use it to map input and stored features into each other. This way, the full information contained...See the full content of this document
Sponsored links
