Subthreshold Current Mode Design of Gabor-type CNN Image Filters

We have reported CNN implementations of 1D and 2D focal plane architectures for filtering images with odd and even symmetric orientation selective filters similar to Gabor filters. These implementations used transistor circuits operating above threshold. While these implementations are fast, they also require a large amount of power. Since many applications may not require extremely fast frame rates, it is reasonable to trade some of the computing speed for reduced power consumption. In this paper, we describe a circuit architecture for Gabor-type image filtering using transistors operating in subthreshold. We exploit the concept of the transistor as a pseudo-conductance or diffuser and use a current, rather than voltage as the primary representation of the signals of interest. This design results more energy efficient computation of the filter responses with similar circuit complexity. Test results from a fabricated 32 pixel 1D array are presented.