Research On Remote Sensing Image Segmentation Based On Encoder-decoder Structured Convolutional Neural Network

Remote sensing image segmentation can use the computer to obtain the information in the image,predict the category of each pixel,and then segment the target from the background.As one of the research contents in the field of computer vision,remote sensing image segmentation can segment the target area from the image,which has important uses in the military and civilian fields.Although image segmentation methods based on deep learning technology have strong segmentation capabilities,remote sensing images have complex scenes,large differences in spectral information,and the characteristics of occlusion,shape,and size of targets,which make it difficult to accurately extract targets by such methods.Features affect the accuracy of segmentation.Taking into account the problems of occlusion and different shapes and sizes of objects in remote sensing images,this thesis focuses on the problem of insufficient segmentation accuracy of remote sensing objects based on the theory of convolutional neural networks and researches on mainstream image segmentation models.The main work of this thesis is as follows:(1)The scene in the remote sensing image is complex,the target is often occluded,and the model is susceptible to interference from background features when extracting target features.In order to obtain accurate segmentation results,this chapter proposes a remote sensing image segmentation model based on attention mechanism.The model adopts an encoding-decoding structure.First,the channel attention mechanism is used to filter the feature information of the encoder.By applying attention weights to the channels of the feature map,the intensity of different channels is adjusted to increase the model's attention to the target feature.Secondly,the position attention mechanism is used to apply position attention to the deep feature map of the encoder,and attention weight is applied to each pixel according to the relationship between different pixels on the feature map to strengthen the feature of the target area.Experimental results show that the model can improve the segmentation accuracy of occluded objects.(2)Remote sensing images involve different scenes,the objects to be segmented usually have different shapes and sizes,which increases the difficulty of the segmentation task.Mainstream image segmentation models based on convolutional neural networks usually use square convolution kernels,which can only extract the features of square receptive fields.The shape of the target in the remote sensing image is changeable,and the square receptive field cannot fit the target shape well,and too many background features may be introduced,which affects the improvement of segmentation accuracy.In order to alleviate this problem,a remote sensing image segmentation model based on adaptive receptive field mechanism is proposed.The model adopts an encoding-decoding structure.First,the features of different aspect ratio receptive fields are extracted through the convolutional layer series combination,and then the attention mechanism is used to weight the features to strengthen the features of the receptive fields that fit the target shape and weaken them.Experimental results show that when the target shape is different,the model can segment the target more accurately,and there are fewer misclassifications and omissions.(3)Remote sensing image segmentation is a pixel-level classification task.By extracting different receptive field features and feature fusion,the model can improve the segmentation accuracy of targets of different sizes and target edges.For this reason,this chapter proposes a remote sensing image segmentation model combining MixConv and SPConv,and the model adopts an encoding-decoding structure.Extracting the features of different receptive fields through convolution kernels of different sizes can cope with the segmentation task of objects of different sizes in remote sensing images.At the same time,by fusing the features of different layers,it helps the model to obtain richer target information,thereby improving the model's segmentation accuracy.Experimental results show that the model can segment targets of different sizes more accurately,and the edges of the targets are clear in the segmentation results.