A heterogeneous SIMD/MIMD PE architecture is proposed to accelerate feedback attention based object recognition. Due to the difficulty of pipelining subsequent stages in an iterative recognition-attention feedback loop, parallelism within the recognition stage is exploited to improve performance. In addition, resource-aware fine-grained task scheduling is performed for high PE utilization, and voltage/frequency throttling based on neuro-fuzzy prediction enables low power real-time recognition on 30fps VGA video.