Research and Implementation of SemanticSegmentation Network Accelerator Based on Vitis AI

Based on Xilinx Vitis AI, this paper carries out network pruning, deep learn- ing processing unit (DPU) customization, and hardware-software co-optimization methods for the semantic segmentation network U-Net. Finally, the semantic seg- mentation accelerator is designed and implemented on the Xilinx ZCU102 develop- ment platform. Under the premise of relatively low accuracy loss, the consumption of hardware resources is reduced, and the complete software and hardware system development of the U-Net network is completed. Experimental results show that the processing frame rate of the U-Net network hardware accelerator can reach 42 fps, verifying the effectiveness of the acceleration scheme for the neural network.