An Energy Efficient Real-Time Object Recognition Processor with Neuro-Fuzzy Controlled Workload-aware Task Pipelining

Abstract

An energy efficient pipelined architecture is proposed for multi-core object recognition processor. The proposed neuro-fuzzy controller and intelligent estimation of the workload of input video stream enable seamless pipelined operation of the 3 object recognition tasks. The neuro-fuzzy controller extracts the fine-grained region-of-interest, and its task pipelining achieves 60.6fps, 5.8x higher performance. The 8.1mJ/frame energy efficiency is obtained by workload aware task scheduling and power management, and it is 6.9x higher efficiency compared to the previous multi-core architectures.