关键词: 深度学习处理单元, ENET网络, 语义分割

Abstract: With the wide application of artificial intelligence (AI) in complex environment such as automatic driving and wearable, an efficient semantic segmentation model has become an important solution object of neural network model. Based on the traditional ENET network model, propose to improve the ENET network. The improved ENET network can use the EeLU activation function hardware mode in DPU to reduce parameters and improve the working performance of DPU. By building the semantic segmentation Vitis-AI architecture platform, the training and learning of constructing quantitative model and model network are completed. A variety of semantic segmentation test results are compared and analyzed. The improved ENET network uses less computing resources to achieve the optimal accuracy. It is deployed on the hardware platform of ZCU106, the performance of the improved ENET network is analyzed and the test results are consistent with the simulation results.

Key words: deeplearningprocessorunit, ENETnetwork, semanticsegmentation