FPGA-Powered Solar Photovoltaic Module Defect Classification: Patch-Wise Reusable Convolutional Neural Network Intellectual Properties for High-Speed Edge Processing

Abstract Modern deep neural network models can achieve high accuracy for computer vision tasks; however, their high computational cost keeps them ashore from deployment on embedded devices. To resolve this, MobileNet was introduced. MobileNet is a lightweight convolutional neural network (CNN) architecture that adopts depthwise separable convolution over the standard convolution to reduce the number of operations and parameters without much loss in accuracy. This article presents the design and implementation of four reusable computing engines for depthwise convolution, pointwise convolution, standard convolution, and batch normalization layers. These engines are designed for low-latency defect classification in solar cells on embedded devices. Our MobileNet model achieved an accuracy of 91% (2 classes), 86% (8 classes), 79% (11 classes), and 86% (12 classes) in classification. Intellectual property (IP) blocks also acquired an optimum execution time of 0.07 ms for pointwise IP, 0.02 ms for convolution IP, 0.13 ms for depthwise IP, and 0.07 ms for batch normalization when deployed on ZYNQ-ZCU104 board.