Research On Road Segmentation Of High Resolution Remote Sensing Images Based On Depth Convolution Neural Network

With the development of aerospace technology and computer technology,remote sensing technology has achieved a trend of vigorous development.How to process massive remote sensing images and obtain effective information from them has become a core issue in the field of remote sensing.In the field of remote sensing,the effective feature segmentation of remote sensing images is the basis of remote sensing technology research,and it is also a hot issue in the field.Therefore,obtaining an effective feature segmentation result map from remote sensing images is the top priority of the development of remote sensing technology.The feature segmentation method based on deep learning can extract positioning information and semantic information from remote sensing images,so as to achieve end-to-end feature segmentation for remote sensing images.Research is based on high-resolution road segmentation tasks using deep learning methods,and uses the Deep Lab V3+ network model to achieve the semantic segmentation of the road segmentation dataset,and optimizes the use of the Deep Lab V3+ network model based on the decoder structure and attention module.The main research work is as follows:（1）Implementation of remote sensing road semantic segmentation method based on deep convolutional neural network.Four models of FCN network model,U-Net network model,Deep Lab V3 network model and Deep Lab V3+ network model are used for training and accuracy evaluation on the Deep Globe Road Extraction dataset.The results show that the Deep Lab V3+ network model performs the best in the Deep Globe Road Extraction dataset.（2）Optimized design and implementation of Deep Lab V3+ network model.Using the idea of feature pyramid and layer-by-layer fusion to redesign its decoder network,optimize the structure of the decoder network,and use the channel attention mechanism and position attention mechanism to optimize the dependency of the feature map before the feature map is fused.The network model can produce a more distinguishable feature representation,and finally the training loss function is optimized to balance the influence of positive samples and negative samples on training.The final experiment proves that the optimization improves the segmentation accuracy of the network model,brings a 1.28% improvement in recognition accuracy.To sum up,using the Deep Lab V3+ network model to achieve the segmentation of the high-resolution road segmentation dataset.The Deep Lab V3+ network model is optimized using the feature pyramid,layer-by-layer fusion and attention mechanism,and then the balance loss function is used to alleviate the problem of the imbalance of positive and negative samples in the dataset,and the experiment proves the effectiveness of the optimization work.