A SoPC architecture for nearest-neighbor based learning and recognition

This paper proposes an efficient and flexible nearest-neighbor-based learning and recognition system on a programmable chip. The nearest neighbor search (NNS), which is a common computational problem in pattern recognition, is to find the minimal distance among all distances between an unknown sample and the complete set of the often high-dimensional reference database. In this paper, a reference optimization learning stage based on NNS is used to reduce the storage space of references and to speed-up the recognition. For efficiently solving the NNS computational problem, we present a specialized hardware implementation with p-component-parallel word-serial pipelined architecture based on the DSP blocks in Stratix FPGA device, used as NNS co-processor. For flexibility, the FPGA embedded processor, Altera's Nios II processor, is used for reference optimization which is the higher level part in the learning stage. Due to the hardware and embedded software cooperation, high-speed learning and recognition based on NNS can be carried out on a programmable FPGA.