| The increasing maturity of remote sensing technology and the development and improvement of remote sensing imaging methods have greatly improved the quantity and quality of remote sensing images.Remote sensing image segmentation is the basis for subsequent image analysis,vegetation and land detection.High-resolution remote sensing images have gradually become an important source of information for surface detection and feature data processing in scientific research.It has the characteristics of rich details,high complexity of features,and many imaging spectrum segments.It is more difficult to segment it.Big.Due to the increase in the number of imaging,while clutter,imaging height,angle,etc.have a greater impact on the quality of remote sensing images,the segmentation effects of traditional threshold segmentation,manual extraction and other methods cannot meet the accuracy needs of subsequent research,and the required labor cost High,how to efficiently and intelligently process remote sensing data has become a technical problem.At present,deep learning technology has made great progress.Among them,neural network algorithms are gradually applied to image processing,machine vision and other aspects,including remote sensing information processing,with their superior performance in automation and intelligence.Aiming at the problem that the remote sensing image has more complex feature information,which makes it more difficult to segment it,the article conducts research and analysis,and the key work includes the following:(1)For the problem of small data set and weak label,the original image and data annotation image are enhanced.Through a certain step of image cutting,the data set can reach a suitable training size,and through changing the contrast of the original image and the labeled image,synchronous image reversal,random rotation and translation transformation,the data set can be enlarged.At the same time,resampling the data set can avoid the training over fitting and uneven distribution of the training set data caused by the small data set with weak label balance to ensure the accuracy of verification set and segmentation effect.(2)In view of the complex information of features(such as road network,water body,mountain area,vegetation artificial building,etc.),the uneven distribution of features,the small scale of some features,the difficult area of classification edge,etc.,an improved symmetric encoding and decoding network structure SegProNet is proposed to realize the end-to-end pixel classification,positioning,encoder part,etc After extracting image features,the decoder can map the classification map of the feature map and map the different classification features learned by the encoder to the pixel space.Using pool index and convolution to fuse semantic information and image features,selectively discarding pool can reduce the loss of image spatial information,and realize the accurate extraction of detailed features.By setting 1×1 convolution at the connection of encoder and decoder,the bottleeck layer is constructed,which can further deepen the network to extract features and reduce the amount of model training parameters,thus improving the networb training speed The coding structure gradually deepens the filter depth to build the end-to-end(semantic segmentation network,and improves the activation function to further improve the network performance.(3)Realize the training of SegNet,U-Net and other classification models,improve the SegProNet activation function,introduce non negative index term to improve the impact of silent neurons on the overall accuracy of the model;make feature refinement after model efficiency analysis,fuse the trained multiple classification learners,sample the classification prediction image samples independently,and realize the casting of classification results In order to improve the generalization ability of the model,the segmentation results of multiple classifiers are integrated and optimized. |
PDF Full Text Request