Abstract: Benefiting from the powerful ability of deep convolutional neural network in feature extraction and semantic understanding, semantic segmentation technology based on deep neural network has gradually become a hot topic in computer vision research. Accurate and efficient semantic segmentation techniques are needed in the fields of unmanned driving, medical images, virtual interaction, augmented reality and so on. Semantic segmentation starts from pixel-level understanding of the image and assigns a separate category label to each pixel. Aiming at the semantic segmentation technology based on deep neural network, according to the differences in technical characteristics, the advantages and disadvantages of existing models are sorted out and analyzed from the aspects of encoder-decoder architecture, multi-scale target fusion, convolution optimization, attention mechanism, traditional-deep combination, and strategy fusion. The current mainstream semantic segmentation methods are compared in the experimental results of public datasets. Finally, the current challenges and future research directions in this field were summarized.