Research And Application Of Image Segmentation Algorithm Based On Deep Neural Network

Image segmentation is a key step from image processing to image analysis.The quality of segmentation directly affects the performance of subsequent image analysis.Remote sensing image target segmentation is one of the important applications in image processing.It can extract targets such as buildings,roads,coastlines,and urban features by processing the entire remote sensing image.It can provide guidance for urban planning,environmental modeling,and natural disaster detection,and has high military and civil value.Remote sensing image has the characteristics of large data scale,small and dense targets,and low color contrast,so there are many difficulties in the research of remote sensing image segmentation.At present,there are many image segmentation methods based on the mainstream depth neural network.However,due to the lack of a large number of labeled images to train deep networks in the field of remote sensing,and the loss of information caused by repeated pooling and convolution steps in the mainstream framework,the model can not get more accurate segmentation results.Based on the above analysis,this paper,on the basis of the existing depth neural network theory,explores and solves the problems of the missed pixel region and blurred target details caused by information loss in the mainstream depth segmentation model,and also unable to effectively cope with the multi-scale of the target in the image.The specific research contents are as follows:(1)Aiming at the problem of information loss caused by pooling and convolution step in deep segmentation model,it can not obtain fine segmentation results.An image segmentation model based on cascaded multi-level features is proposed.Firstly,the last layer convolution feature of the first two convolution stages and all convolution layer features of the last three convolution stages in the encoder are selected,and the features of the latter three stages are added pixel by pixel through skip connection.Secondly,all the convolution layer features of the last three convolution stages in the decoder are selected to fuse pixel by pixel.Finally,the above multi-level features are cascaded in the way of channel splicing,and then sent to the new convolution layer to learn and make category prediction.Experiments on Caltech-UCSD Birds200 and ISPRS Vaihingen datasets and comparisons with other depth segmentation models.The results show that cascading feature maps at different levels enables the model to take full advantage of the intermediate layer features extracted by the deep network,and the segmentation targets are more complete,continuous,with fewer misclassifications and Leakage phenomenon.(2)Aiming at the problem of information loss caused by pooling and convolution step in depth model,and different target scale,complicated illumination and occlusion in remote sensing image,a remote sensing image segmentation model combining complete residual connection and feature fusion is proposed.The residual unit is used by the corresponding stages in the encoder and decoder.At the same time,the feature information of several convolutional layers in each convolution stage of the encoder is extracted,and the feature pyramid module is used to extract multi-scale features from the feature map of the last convolution stage of the encoder.Finally,the above feature information is fused into the corresponding layer of the decoder by pixel by pixel addition.Experiments were carried out on ISPRS Vaihingen and Road Detection datasets and unlabeled images.The results show that the proposed model outperforms the current advanced image semantics segmentation model,and the segmentation goals are more complete,continuous,with fewer misclassifications and Leakages.The results of road segmentation from different remote sensing images are better than those of comparative models.(3)Aiming at the problem of poor spatial continuity of pixels in the result image of the mainstream depth segmentation model,and the computational complexity and high cost of the existing post-processing methods.Based on the existing deep network segmentation model,a cGAN image segmentation model combining multi-scale context information is proposed.It consists of a generator and a discriminator.The generator is an improved SegNet model extracting the multi-scale features of the end-pooling feature map of the encoder by using different rate of dilated convolution and fusing them with global features to achieve the purpose of extracting multi-scale context information.The discriminator is a two-class convolutional neural network whose input is the combination of the generated image or the real marker map and the original image.It can realize the purpose of judging whether the input comes from the generated image or the real label image.By comparing with the comparison model on the Road Detection dataset,the results show that the segmentation results of the model are better than the comparison model in terms of integrity,detail and spatial continuity of pixels.