Improved U-NET Semantic Segmentation Network

Abstract

In order to enable neural networks to better process images, this paper proposed a network with residual network and attention mechanism (AM). The residual network solves the shortcomings of insufficient depth of U-NET network to a certain extent. The attention mechanism can adaptively fuse processed context information and place different attentions for different objects and shapes, instead of simply aggregating all processed information. The method proposed in this paper is compared with the network using the residual network alone and the network using the attention mechanism alone. The experimental results show that the network proposed in this paper performs better.