An Improved Semantic Segmentation Model Based on FCN with Channel Attention and Feature Fusion

This paper proposes an improved semantic segmentation model based on Fully Convolutional Network(FCN). Firstly, this paper integrates the channel attention module into the backbone of FCN to identify the channels that are more important for semantic segmentation. Secondly, this paper adds a multi-scale feature fusion module to the network to integrate multi-scale feature information. Thirdly, this paper replaces some of standard convolutions in the backbone network with dilated convolutions to increase the receptive field. Experimental results demonstrate that after adding the two modules, the pixel accuracy can reach 75.8%, and the mIOU value has increased by 2.8% compared to the original FCN model. Moreover, utilizing pre-trained weights during the training process can greatly enhance the performance of the model.