Semantic Segmentation Model Based on Adaptive Fusion and Attention Refinement

Aiming at the insufficient use of context information and loss of detail information of the existing semantic segmentation, a model based on adaptive fusion and attention refinement is proposed. The model introduces an adaptive fusion module in the process of coding, and solves the insufficient use of context information by fusing each feature map according to the corresponding weight. An attention thinning module is designed in the process of decoding, so that the low-order features and high-order features can guide and optimize each other to solve the loss of detail information. The experimental results show that the average intersection union ratio of the model on PASCAL VOC 2012 dataset reaches 83.7%, which is 1.1% higher than the semantic segmentation model based on encoding and decoding. The average intersection union ratio of 81.7% is obtained on cityscapes dataset, which further verifies the generalization of the model.