Semantic Segmentation Of Remote Sensing Images Based On Deep Learning

As an important part of computer vision,image semantic segmentation can identify both the location of the target and the category of target,which has important practical significance in remote sensing image processing.Based on the traditional semantic segmentation algorithm and deep convolutional neural network,an improved semantic segmentation model,ELGNet,is proposed.The main work of this paper is as follows:(1)This paper analyzes the negative impact of the inequality of sample categories on model performance in detail and reduces it from two aspects of data preprocessing and loss function selection to improve network accuracy.(2)On the basis of EfficientNet,this paper proposes improvement measures of network model.For the encoding part,the global and local interaction modules are added to slow down the information loss during the network long-distance convolution,thus improving the segmentation accuracy.For the decoding part,the multi-scale information of the encoding part is added,which plays a positive role in restoring the spatial resolution of the image.In addition,an up-sampling experiment is carried out to select an appropriate deconvolution method for the network.(3)The deep learning model,ELGNet,is applied to the semantic segmentation of two remote sensing image data sets and it is compared with other models.The results show that BCE as the loss function is better than Focal loss in the data set used in this paper.The interpolation method for the upper sampling is better than the transpose convolution,and the bicubic interpolation is slightly better than the bilinear interpolation.The optimized model in this paper can improve mIoU on both data sets by about 8% compared with the original model.In the case of similar accuracy,the training speed of the model used in this paper is more than twice as fast as that of FCN 8s.