On Edge Image Processing Acceleration with Low Power Neuro-Memristive Segmented Crossbar Array Architecture
Nikolaos Vasileiadis, Vasileios Ntinas, Panagiotis Karakolis, Panagiotis Dimitrakis and Georgios Ch. Sirakoulis
Computational acceleration for image processing tasks on the edge is becoming increasingly important for many applications. This work presents a new neuro-inspired architecture which incorporates in-memory computing properties for image processing complex computational tasks in addition to analog memory. The proposed architecture was based on segmented crossbar topology, which, as it turns out, reduces many phenomena that affect the performance on such systems. The extended architectural capabilities of this structure were also tested in a systematic analysis that was performed on a novel depth map extraction application from a single defocused image. All results were validated through Spice simulations using a novel moving barrier memristor model.
Keywords: Edge computing, IoT, memristor, neuromorphic accelerator, inmemory computing, segmented crossbar, convolution engine, depth map