Invariant object recognition with a neurobiological slant

Position-, scale- and rotation-invariant (PSRI) object recognition (OR) aims at recognising an object independently of its position, scale and orientation. Those who have applied artificial neural nets (ANNs) to this subject have loosely followed two different approaches. The first is to remove the variance within the input data in a non-ANN preprocessing stage, and then present this `normalised´ information to the neural net. The second approach is to take into account the variance in the input data within the architecture of the net. We have developed an unsupervised PSRI prototype InVariant OR system (IvOR) that falls into the first category, but with preprocessing stages with neurobiological inspirations. An important aspect is the automated recognition of rogue features on cast objects. IvOR 1 has had limited success in differentiating such features. This paper gives a brief description of IvOR 1 and then outlines a modified system, IvOR 2, where the approach taken employs a structure based on the hypercolumn model of the striate cortex as a feature extractor. A novel technique to elicit relevant features from an image and then ascertain the correct relationship between the features has been implemented. This technique creates an input to a learning system comprising a modular arrangement of ART nets. A series of ART2a nets categorise and learn object feature representations. The final categorisation is done by an overseeing ART1 net that takes input from the feature representing ART2a modules. Finally, the current state of the work is discussed and future developments are outlined