A Bio-Inspired Reconfigurable Architecture for Local Binary Descriptors

Limited resources of embedded devices and increased real time constraints have raised interest to the binary description methods over the floating-point ones such as SIFT and SURF in computer vision applications. Although many software applications of the binary descriptors are developed, there are few studies on their hardware implementations in the literature. Despite the fact that direct hardware implementation of the algorithms can enhance the performance of calculations, flexible architectures which are able to support new approaches remain still an issue. At this point, mimicking biological neural network which has naturally capable of extracting features can offer different insights to the problem. In this paper we propose a reconfigurable architecture and its’ hardware implementation for local binary description applications, inspiring from the biological neural network structures. Specifically, this architecture is based on the Cellular Neural Network (CNN), with two different types of cells which show either inhibitory or excitatory behaviour. Imposed restrictions on classical CNN architecture lead a multiplier-less area-efficient realization, while keeping the necessary flexibility. This network can be configured as an accelerator unit of the local binary descriptors in embedded vision applications.