Multitask Semantic Segmentation Network Using Adaptive Multiscale Feature Fusion

Abstract

A multi-task semantic segmentation network architecture based on adaptive multi-scale feature fusion is proposed, which improves segmentation target edge details and small-scale target segmentation accuracy by combining boundary detection tasks and semantic segmentation tasks. The critical component of the architecture is the adaptive multi-scale feature fusion module, which can adaptively fuse the semantic feature information and boundary feature information of different scales, extract semantic features that contain more spatial data, and reduce the loss of spatial information of small-scale targets, thereby enhancing the network's ability to learn small-scale target features and boundary features. Experiments show that our designed network architecture can improve the segmentation accuracy of small-scale objects and optimize the edge details of segmented objects.