Semantic segmentation based on DeeplabV3+ with multiple fusions of low-level features

Abstract

For the problem that DeeplabV3+ semantic segmentation model uses downsampling operation several times in the encoder process, a large amount of object boundary information is lost, resulting in inaccurate segmentation at the object boundary location, this paper proposes a new semantic segmentation model that fuses low-level features several times based on DeeplabV3+ algorithm. Firstly, we use the encoding module of DeeplabV3+ algorithm to extract the high-level semantic information of the object; then we use the method of this paper to extract the low-level features several times to obtain the detailed information of the object boundary; finally, we fuse the high-level semantic information and the detailed information of the object to obtain the optimized semantic segmentation result. The experimental results on the current open source dataset PASCAL VOC 2012 show that compared with the DeeplabV3+ algorithm, the algorithm model proposed in this paper has better semantic segmentation results by fusing the low-level features multiple times, especially at the boundary position of the object, with a pixel accuracy of 94.1% and a mean intersection over union of 77.5%. The overall performance of the algorithm model proposed in this paper has reached the current leading level.