Attention Mechanism Driven YOLOv3 on FPGA Acceleration for Efficient Vision Based Defect Inspection

Abstract

In this study, an efficient vision-based industry defect inspection system using attention mechanism driven YOLOv3 on FPGA acceleration is proposed. First, an attention mechanism is employed to improve YOLOv3 for the target defect inspection application. Image preprocessing named CZS (Cut, Zoom, and Splice) operation is used to reconstruct product images for selectively concentrating on the pre-defined detection regions. Then we optimize the backbone network of YOLOv3 according to defect size in images. Second, we use the PYNQ-Z2 FPGA board to deploy the proposed defect inspection system. The optimized YOLOv3 is deployed on the programmable logic through Xilinx DNNDK, which is a low-latency, low-cost, and low-power consumption hardware platform for industrial defect inspection. Experimental results showed that the achieved defect inspection accuracy was 99.2% with a processing speed of 1.54 FPS.