Multi Path Feature Fusion Network Based On Deep Learning For Semantic Image Segmentation

Semantic image segmentation is one of the key issues in computer vision and the goal is to predict the category label of each pixel in an image.Image semantic segmentation plays an important role in many fields.As the basic algorithm of these fields,the performance of semantic segmentation determines the effectiveness of subsequent algorithms.Therefore,it is of great significance to improve the performance of semantic segmentation algorithms.This paper takes the deep learning technology as the starting point,and proposes the method of multi-path feature fusion to improve the performance of semantic image segmentation.After carefully reading many papers and researching related works,this paper summarizes the existing methods and problems to be solved in semantic segmentation.The proposed innovations and improvement programs are as follows:(1)In order to extract and fuse the multi-scale detail features,in the feature extraction stage(Encoder),a multi-path feature fusion module is proposed,which allows images to pass through different paths and output detailed features with different scales,and then these features are fused to generate more representative features.The deep convolutional network is then built using the multi-path feature fusion module.(2)In this paper,the global average pooling layer is added at the end of the feature extraction network to obtain global features,and the global features and local features are merged.In the decoder stage,a stacked pooling block is proposed,and combined with the features from convolutional layers to obtain complementary image discriminant features for image semantic segmentation,so as to further improve the precision of semantic segmentation.In order to verify the effectiveness of the proposed method,comparative experiments was conducted on two public data sets,CamVid and CityScapes.Compared with FCN,mIoU increased by 6.4%and 2.1%respectively.The experimental results show that each module of the network contributes to the performance improvement of the overall network,and the performance of our network is better than many recently proposed networks.