Research on ZYNQ neural network acceleration method for aluminum surface microdefects

Abstract

Convolutional Neural Networks (CNN) are an important means of detection of microdefects on the aluminum surface, and the high complexity and computing power requirements of the CNN model lead to difficulties in deploying them on edge computing platforms as the detection accuracy continues to improve. We have studied a lightweight acceleration method for detecting microdefects on aluminum surfaces on the Zynq-7000 All Programmable SoC (ZYNQ) platform. A lightweight aluminum surface defect detection network (LADFastDet) and high-performance accelerators based on ZYNQ are designed to meet the requirements of precision and speed under limited resources. In the LADFastDet structure, a lightweight inverted residual block is designed by combining depthwise convolution, inverted residual block, and inverted bottleneck. A multiscale feature fusion structure is designed to effectively improve the detection accuracy of LADFastDet, especially small target defects. We design accelerators on ZYNQ through optimization methods such as loop optimization strategy, ping-pong buffering, and multichannel and multiple interfaces data reading and writing to reduce data access latency and thus improve the computing speed. The experimental results show that the LADFastDet model has a mAP of 97.51%, the inference time of the accelerators for a single image is 42.57 ms, and a power consumption of 2.15 W, which achieves a throughput of 24.9 GOPS and an energy efficiency of 11.58 GOPS/W.