A Real-time Semantic Segmentation Method Based on Multi-level Feature Fusion

Abstract

The performance improvement for real-time segmentation networks is generally to accelerate the segmentation speed of the model at the cost of computational cost, ignoring the problem of semantic inconsistency of neighborhood features, which causes the accuracy of segmented images to decrease. Therefore, it is crucial to take into account the segmentation efficiency while ensuring the accuracy of model segmentation. In this paper, a lightweight model based on Multi-level Feature Fusion Semantic Segmentation Network (MLFFNet) is proposed, and the network as a whole adopts a two-branch structure to differentiate different types of features. The model obtained 81.4 FPS forward inference speed and 71.3% segmentation accuracy on the Cityscapes dataset, which is capable of real-time semantic segmentation tasks and proposes a new idea for the semantic segmentation problem in a complex context.