A semantic segmentation method for unstructured complex environments

Abstract

This invention discloses a semantic segmentation method for unstructured complex environments, comprising the following steps: 1) Modifying network feature fusion by replacing pooling with dilated convolution feature fusion to reduce the impact of pooling on network feature extraction, and using feature fusion at different scales to expand the network's receptive field and improve the overall network's segmentation accuracy. 2) Establishing a coordinate attention mechanism module based on small object categories to improve the network's semantic segmentation accuracy for different complex roads and small objects. 3) Constructing a novel dual-weight loss function to address the imbalanced sample categories in unstructured roads by assigning different weights to image sample categories to further improve the network's segmentation accuracy for small object categories in images. 4) Using a pre-trained network model to extract features from the input image and feeding them into the improved feature fusion-attention ICNet for prediction.

Images