Research On Semantic Segmentation Method Of Remote Sensing Image Based On Deep Convolutional Neural Network

With the rapid development of remote sensing technology,the acquisition of remote sensing images is no longer difficult,and the quality and quantity of remote sensing data have increased exponentially.Semantic segmentation of high-resolution remote sensing images has become an important technique of earth observation by processing and utilizing remote sensing data.In recent years,deep learning has achieved excellent achievement in many computer vision tasks.Image segmentation using deep convolutional neural networks has also become a hotspot in academia.Compared with traditional image segmentation algorithms that manually extract features,the method based on deep convolutional neural networks has stronger feature extraction capabilities and has shown superior performance in ordinary natural image segmentation tasks.However,there are huge differences between remote sensing images and ordinary natural images.The land cover categories in remote sensing images are highly imbalanced,large targets dominate the segmentation task,and small targets are often suppressed.Moreover,there are multiple scale targets in high-resolution remote sensing images,and being able to fully identify targets at various scales has always been a difficult point in the field of remote sensing image segmentation.Aiming at the above problems,this paper analyzes the characteristics of high-resolution remote sensing images in depth,researches on the problem of class imbalance and multi-scale target segmentation in high resolution remote sensing images,and achieves a good segmentation effect.The main research work of this paper is as follows:(1)For the problem of category imbalance in the semantic segmentation task of highresolution remote sensing images,this paper proposes a focal loss function based on the median frequency balance.The loss function first calculates the ratio of the median value of the training sample category frequency to the target category frequency.Then use this ratio to weight the focal loss function,suppress the dominant role of the large targets during training,and enhance the classification effect of small targets.(2)In order to improve the segmentation effect of high-resolution remote sensing images,this paper proposes a long-short connection network that combines multi-layer features.The network has a symmetric encoding and decoding structure.By constructing a short connection module that connects the convolutional layer,the features are extracted in the down-sampling stage,and partially reconstructed at the up-sampling stage,and the gradient descent problem that occurs as the neural network deepens is solved.The feature image of the corresponding part of the encoding network and the decoding network is fused by using the long connection operation,and the detailed information of the shallow feature map and the abstract information of the deep feature map are fused.Compared with U-Net,LSCNet has a better overall classification effect,especially for the small target category.(3)In order to further improve the effect of image segmentation,this paper conducts research on multi-scale segmentation models,and proposes a multi-scale parallel shortcut network.The network design a parallel shortcut module composed of multi-branch dilated convolutions in parallel.The convolution of the different branches has different dilated rates,and the number of branches included in the module is adjusted according to the needs of the network.Dilated convolution with different dilated rates can capture the characteristics of receptive fields at different scales while expanding the receptive fields.At the same time,an improved pyramid pooling module is introduced in the network to aggregate context information of different scales,improve the network’s ability to use global context information,and further improve the multi-scale segmentation performance of the network.