Stripe pooling and densely connected DeepLabv3 plus efficient semantic segmentation

Abstract

Currently, DeepLab is unable to utilize multiscale feature information at multiple levels, and there are often problems such as blurred segmentation boundaries, unclear detail extraction, and incorrect segmentation.This article optimizes the DeepLabv3 plus model The backbone network has been converted to a lightweight MobileNetV2 network. In Atrous Spatial Pyramid Pooling (ASPP), stripe pooling has been used to replace global average pooling, and the original hole ratio combination of 6, 12, and 18 has been changed to 3, 7, 9, and 17. A branch with R=3 has been added, as well as the use of dense connections. The improved ASPP has the advantage of higher acceptability. The experiment shows that the average intersection ratio of the improved DeepLabv3 plus model on the dataset is 69.71%, and the average pixel accuracy is 79.45%. Compared with the original network model, the improved average intersection ratio is increased by 3.2%. Using the above improved methods has improved the performance of DeepLabv3 plus, enabling more detailed information to be obtained, and improving the resolution of the model.