Abstract
Biological vision systems use a complex hierarchical structure of features to represent the world. Lower level features (retina to V1) include intensity gradients and changes, oriented edges, motion, color, etc. For mammals, object classification performance greatly depends on both the base feature set and the hierarchical structure that combines them. Object classification becomes even more difficult in changing environments where the useful features for detecting and classifying a given object may change. We propose a model, based on the mammalian visual system, which learns a hierarchical structure to combine a set of base features which are used to detect and classify a specific object. The model continually adapts to handle changing environments and involves a two-stage process: potential objects of interest are identified in the first stage and passed onto a second which provides a behavioral reinforcement signal. The first stage uses parallel background subtraction models, one for each feature, combined via a weighted voting mechanism and produces a ranked list of potential objects. Feedback from the second stage adjusts the weights from each model using a modified instar learning law to lower both false positive and false negative rates. The result is a combination of basis features that are useful for detecting a specific object. We tested our model with an object identification task in a noisy environment. Simulation results demonstrate that learning an appropriate feature basis increases performance (true positive rate) three-fold (0.1 to 0.34) by selecting a set of 2/8 total features and seven-fold (0.1 to 0.7) by selecting a set of 12/134 total features. Learning occurred over tens of frames, enabling quick adaptation to changing environments. These results demonstrate that our model is able to learn to appropriately combine features and may provide insight into the biological structure used by the human visual system.
Meeting abstract presented at VSS 2013