Research On Semantic Segmentation Algorithm Of Remote Sensing Image Based On Improved Fully Convolutional Network

Semantic segmentation of images is an important research content in the field of computer vision,especially the semantic segmentation of remote sensing images,which plays an important role in geographic mapping,geological hazard assessment and other aspects.With the continuous development of deep learning,including the rise of convolution algorithm of neural network is to promote the rapid development of the semantic segmentation,compared with the traditional semantic segmentation method,the semantic segmentation method based on convolution neural network with its intelligent,accurate and efficient advantages have been widely used,the basic of image semantic segmentation using convolution algorithm of neural network,but also exist in solving some specific problems of insufficient.In terms of semantic segmentation of remote sensing images,the problem of unbalanced distribution of object categories exists in the remote sensing image data set itself,which will lead to the network prefering to learn the big target or the big background and ignoring the small target,thus leading to low segmentation accuracy.Therefore,the research work of this paper is carried out on the basis of the above.The research work and contributions of this article are as follows:（1）Starting from the basic convolutional neural network,the basic structure and algorithm principles are studied,laying a foundation for the subsequent improvement of the full convolutional network,and further researching the full convolutional network algorithm flow for semantic segmentation,and analyzing the imbalance of dataset categories in current semantic segmentation leads to low segmentation accuracy and poor segmentation effect on small targets,which points out the direction for subsequent research work.（2）Improve the network structure of the full convolutional network,analyse the influence of the attention mechanism on the full convolutional network,study the method of recovering the information lost in the coding stage,and use pyramid pooling as a multi-scale feature extraction method,and the residual attention mechanism combined with the fully convolutional network encoder-decoder structure,an end-to-end multi-scale attention network PAU-Net is designed.The experimental verification is carried out on the remote sensing image dataset.The experimental results show that it is compared with the original basic network U-Net.Compared with the two indexes,the mean pixel accuracy of segmentation and mean intersection and union are increased by 5.76%and 4.15%respectively,which optimizes the segmentation effect of small targets.In addition,the generalization experiment results on the public dataset also prove the effectiveness of the improved method.（3）Aiming at the improvement of the loss function of the full convolutional network,the method of weighting the category loss is studied,a loss function L_classbased on the category distribution is proposed,and its effectiveness is verified by experiments.The experimental results show that compared to the original network that only uses a single cross-entropy loss function,the category distribution loss function constructed in this paper also improves the mean pixel accuracy of segmentation and the mean intersection and union of the two indicators,which proved that the loss function proposed in this paper can optimize the network training and reduce the impact of the unbalanced distribution of categories in the dataset.