Research Of Remote Sensing Image Semantic Segmentation Based On Convolutional Neural Networks And Conditional Random Fields

With the increase of orbiting satellites and other aircraft,humans have obtained many remote sensing images,which contain rich surface information.If the information in such images can be fully utilized,then land cover research,animal and plant ecology Protection,the construction of three-dimensional maps,etc.will play a big role in promoting.Semantic segmentation of remote sensing images is the basis for realizing the use of surface information.The current semantic segmentation of remote sensing images faces some problems:(1)Although there are many remote sensing images acquired,the resolution is high and can be used directly are less;(2)Remote sensing images acquire real surface conditions,which means The different objects in the image are interlaced,and the information is too complicated.At the same time,the surface objects are easily affected by the illumination,resulting in different views of the same kind of objects,which makes it difficult to design accurate semantic segmentation algorithms.In view of the above problems,this paper researchs the semantic segmentation of remote sensing images based on the knowledge of convolutional neural networks and conditional random f elds.The main work is as follows:(1)Based on the convolutional neural networks structure,this paper designs a two-channel image feature extraction networks,and combines the ResNetl8 networks pre-training fine-tuning model to prevent the over-fitting phenomenon caused by insufficient training data.Finally,the design completes a multi-scale feature extraction and fusion of deep convolutional neural networks to learn shallow features and deep information of remote sensing images.(2)Due to the downsampling effect of the pooling,the convolutional neural networks will cause some loss of feature information when extracting the feature of the image target.For this reason,this paper adds a band to the designed dual-channel networks structure.The maximum pooling method with position index achieves the purpose of better preservation of image target position information,and achieves better semantic segmentation effect.At the same time,the batch normalization layer is added to further optimize the networks structure,and the convergence speed of the neural networks is accelerated while ensuring the networks training process is stable.Based on the design of these methods,a new deep convolutional networks is constructed to research the semantic segmentation of remote sensing images.(3)On the basis of the above-mentioned deep convolutional networks,in order to better improve the target edge details in the semantic segmentation image and fully consider the consistency of similar pixel annotation,this paper designs a cyclic condition random fields module in iterative solution.Combine it with a deep convolutional networks to more accurately complete the semantic segmentation task of remote sensing images.Based on the above model algorithm,this paper uses the remote sensing image dataset to verify the experimental results.The model algorithm is qualitatively and quantitatively compared with the excellent semantic segmentation algorithm in the research field.The comparison algorithm includes FCN-8s,Unet and SegNet model algorithms.Through experimental comparison,it is concluded that the fully convolutional neural networks model designed in this paper is more effective than other model algorithms.,and the addition of the cyclic conditional random fields module model further improves the accuracy.The algorithm finally achieved an accuracy of 93.36%,Kappa consistency reached 0.8905,and the training time was shortened.Compared with other methods,the algorithm has obvious advantages and achieved good semantic segmentation.