Semantic Segmentation Of Small Targets In Remote Sensing Images Based On Deep Convolutional Neural Network

Semantic segmentation of small targets in remote sensing images has always been a difficult task,and the most representative one is road segmentation.And with the con-tinuous progress in the field of remote sensing,road segmentation of large-size remote sensing images has long become a key breakthrough research direction.As the most basic transportation facility,roads are the foundation of national construction,urban planning,military operations,and economic growth.Currently,most researches are based on tradi-tional machine learning algorithms.Because the images are full of discernible impervious surfaces such as roads and buildings,they are complex and computationally intensive.Al-though this task has been extensively studied in the past few years,it is still a challenging task due to its special characteristics.This thesis reviews the development of road segmentation technology of remote sens-ing image,and proposes a semantic segmentation neural network based on deep learning.This thesis will carry out the research work of this thesis from the following four aspects:(1)Research on the encoder-decoder layer based on U-Net.The network uses a pre-trained neural network to decode image features and introduces a dense expansion layer in the middle part.The architecture of the network is similar to U-Net,that is,the stacking method of the convolutional layer for encoding and decoding has a U-shaped structure in space.The U-Net architecture is effective in enhancing computing performance and saving memory consumption.(2)Research on expanded convolutional layer based on dilated convolution.This layer uses short connection and dilated convolution method to expand the convolution layer.The dense expansion layer constructed by the dilated convolution expands the re-ceptive field of the network and reduces the loss of image feature information,so as to complete the road segmentation of remote sensing images.(3)Research on balanced loss layer based on Dice Loss.Added multi-scale Dice loss to improve the loss function.(4)Research on the problem of reconnecting broken road segments based on proba-bility regularization algorithm.Because of the algorithm results of this network,broken road segments often appear.Therefore,this thesis proposes a probabilistic normalized connection road algorithm to reconnect broken road segments.After exploring the theoretical knowledge,this thesis tests the performance of the network on the DeepGlobe road extraction data set.The specificity,sensitivity,accuracy,recall score,F1 index,kappa coefficient and IoU are used to evaluate the performance of the model,and the comparative test of each module is carried out by the method of controlling variables.The results show that the modules in the network of this article have a positive effect on the model.