关键词: 深度学习, 变分自编码器, 边缘优化

Abstract: Image semantic segmentation is a pixel-level image classification method, which plays an important role in un-manned driving, robot vision and other fields. Aiming at the problems of inaccurate edge segmentation and easy loss of small target details in the process of semantic segmentation, an image semantic segmentation algorithm based on deep learning and superpixel segmentation was proposed. The image features were learned by using variational auto-encoder, and rough full- resolution semantic labels were predicted by multi-decoder structure. At the same time, the rough semantic segmentation results are further optimized on the basis of superpixel image segmentation, and finally verified on COCO2017 datasets. After experimental analysis, the proposed algorithm achieved a 4.4% improvement in global pixel accuracy（GA）compared to the pre-fused approach. Additionally, the average intersection over union（MIoU）increased by 8.2% compared to the pre-fused method, resulting in a 24.3% enhancement in segmentation precision over the SegNet network. The proposed algorithm achieves good results in both the segmentation accuracy and the segmentation speed, and achieves a good balance between real-time and accuracy.

Key words: deep learning, variational auto-encoder, edge optimization