In industrial production, intelligent detection of product defects is crucial. Field-programmable gate arrays ( FPGAs) are embedded devices with features such as high arithmetic power and low power consumption that enable small convolutional neural networks to be deployed in them. In this paper, a set of improved YOLOv2 target detection algorithm is designed based on Xilinx Zynq series FPGAs, and a reordering layer is added to the model framework to complete the detection of surface defects on aluminum sheets by parallel computing processing of the slice map before reorganisation. The algorithm is designed at a high level ( HLS), then RTL converted and IP cores are packaged and imported into the project to complete the SoC design. Generate bitstream files through comprehensive layout and wiring, import them into PYNQ images, and complete industrial defect detection on the surface of aluminum sheets. The experimental results show that this system can accurately detect defects and reduce power consumption to 2. 494 W.