Semantic segmentation based on semantic edge optimization

Semantic segmentation based on deep learning is to extract semantic features by convolution, and then classify each pixel. In the process of feature extraction, the subsampled operation used to expand the receptive field will cause a large amount of detail information loss, resulting in the loss of small-scale objects in the image and blurred segmentation edge. In order to solve this problem, this paper proposes a segmentation model of semantic edge optimization, which uses the end-to-end semantic edge detection network to learn the semantic edge features of the image, and then fuses the semantic edge features with the semantic segmentation features, so that more edge information can be retained in the final segmentation image. On the camvid and cityscape datasets, our optimization model improve over original semantic segmentation model by 1.4% and 1.5% in terms of mean IoU.