Energy-Efficient VLSI Architecture for Industrial Image Processing Using CNN

Abstract

The automated inspection systems in the working plant are new and have to be implemented very quickly; hence, strong and efficient with regards to energy consumption hardware components are needed. The present article is a VLSI architecture of industry picture processing with Convolutional Neural Networks (CNNs) that consumes less energy. The design presented, includes optimised processing elements, low power memory access techniques and parallel computing techniques with a target of high throughput, and low power consumption. This is implemented in a 45 nm CMOS technology that operates on a 500 MHz clock frequency and uses as low as 120 MW of power. The processing throughput of this system is 60 frames/second (fps) with a high-resolution image (1024 x 1024 pixels) which is adequate to achieve a real-time processing in industry such as defects detection and quality control. The proposed architecture requires 35 percent less power, as well as has 28 times more computing capability than the traditional architectures. It is also able to enhance memory utilization by a quarter of data reuse and compression techniques. Defect classification has been demonstrated to be accurate 96.8% in an experiment. The proposed architecture is a cost effective and scaling system to implement next generation smart manufacturing devices which require a high-speed low power image processing system.