Improving the semantic segmentation algorithm of DeepLabv3+

Abstract

There is a conflict between model complexity and segmentation accuracy in existing semantic segmentation algorithms. Based on this, two aspects are considered in terms of the number of parameters and segmentation accuracy. The existing deeplabv3+ algorithm is improved, and the lightweight backbone network Mobilenetv3 is selected for the backbone network to implement the extraction of image features in order to reduce the number of parameters. Putting the coding stage atrous spatial pyramid pooling (ASPP) Replace withDense atrou spatial pyramid pooling (DenseASPP). Compared to ASPP, DenseASPP has a denser pyramidal structure of features and a larger sensory field.At the same time, DenseASPP and SP are concatenated to further enhance the network’s ability to extract context. The MSCA attention mechanism and NAM channel attention are added in the encoding stage and decoding stage respectively to further improve the segmentation accuracy. The average cross-merge ratio on the VOC dataset is 72.46% and the number of parameters is only 7.7 M. Compared with existing pairwise semantic segmentation algorithms, the algorithm achieves a good balance between segmentation accuracy and segmentation speed.