Low power CNN Accelerator for mobile lensless imaging system

This paper presents a CNN accelerator structure for lensless microfluidic acquisition of biological image processing. We propose a low power optimized convolution circuit which utilized the sparse characteristics of lensless cell image acquisition system. And a reconfigurable dual-register group structure is also developed for improving the CNN processing efficiency. The CNN accelerator structure can accelerate various cell image analysis algorithm, including image edge segmentation and super-resolution reconstruction. The circuit is implemented with UMC 110nm process, the circuit area is 2.66 mm
²
and the maximal frequency is 300 MHz.