A novel large kernel two-branch semantic segmentation network

For semantic branching in a two-branch network structure, it is crucial to quickly improve the feeling field, in addition, the feature fusion interaction of two-branching needs to take into account the structural and multi-scale information fusion of features under different feeling fields. In this paper, in order to cope with this challenge, we propose to accumulate the receptive field in dense convolution using separable large kernel convolution through an adaptive semantic extraction module, and design a selectable large kernel attention module for the problem of mismatch of receptive field of feature information. Finally, this paper designs a multiscale feature extraction module that fuses local and global information, and uses large kernel dense separable convolution in the module with the ability to extract high-dimensional semantics serially and capture multiscale in parallel.