Associative Memory based on clustered Neural Networks: Improved model and architecture for Oriented Edge Detection

Abstract

Associative Memories (AM) are storage devices that allow addressing content from part of it, in opposition of classical index-based memories. This property makes them promising candidates for various search challenges including pattern detection in images. Clustered based Neural Networks (CbNN) allow efficient design of AM by providing fast pattern retrieval, especially when implemented in hardware. In particular, they can be used to store and next quickly identify oriented edges in images. However, current models of CbNN only provide good performances when facing erasures in the inputs. This paper introduces several improvements to the CbNN model in order to cope with intrusion and additive noises. Namely, we change the initialization of neurons to account for precise information depending on Euclidean distance. We also update the activation rules accordingly, resulting in an efficient handling of various types of input noise. To complete this paper, associated hardware architectures are presented along with the proposed computation models and those are compared with the existing CbNN implementation. Synthesis results show that among them, several divide the cost of that implementation by 3 while increasing the maximal frequency by 25%.