Research On Convolutional Neural Network Based Satellite Remote Sensing Image Object Recognition Algorirthm

Remote sensing is a comprehensive technology for Earth observation,which is widely used in many fields such as environmental monitoring,urban planning,and resource survey.With the development of satellite remote sensing technology,its data volume and information volume are showing a rapid growth trend.The increase of information volume makes the details of remote sensing images more abundant.Therefore,the processing of remote sensing images has become more complicated.How to identify objects in massive multi-dimensional high-resolution remote sensing images has become a hot topic in this field.In recent years,the convolutional neural network has made great breakthroughs in the field of image processing.This method has automated feature extraction and feature representation capabilities,which greatly improves the accuracy of object recognition.However,this method still has problems such as insufficient feature expression ability and insufficient decision information.The lack of feature expression ability is manifested in both high-level and low-level aspects: the information represented by high-level features is more abstract and can only express coarser overall information;the lower layer information represented by the features is more specific and cannot effectively summarize the characteristics of various types of objects.The problem of insufficient decision information is that the receptive fields of the feature maps generated at different stages are different.The global information and context information obtained by different receptive fields are highly correlated,while the general method only considers the final features.Therefore,this paper proposes two types of full convolutional neural network models based on spatial pyramid pooling and encoder decoder,and uses the multi-dimensional image information to complete the unified multi-object recognition task.The main research results are as follows:For the problem of insufficient feature expression ability,the model based on spatial pyramid pooling method is designed and implemented,which mainly strengthens the high-level features.The model uses multi-scale global average pooling to process the feature map,generate a variety of feature maps with different receptive fields,and obtain more global information through the fusion operation,thereby enhancing the expression of high-level information.The experimental results show that the model improves the overall recognition accuracy by two percentage points,effectively improving the expressive ability of high-level features.For the problem of insufficient feature expression ability,the model based on multi-path reinforcement method is designed and implemented,and features of multiple levels are strengthened.The low-level information is provided by using an encoder-decoder structure,and the multi-path enhancement method is used to perform convolution,pooling,and fusion enhancement of multiple levels of feature maps to make it more expressive.In the experiment,a high recognition rate of small objects was obtained,which effectively solved the problem of insufficient expression ability of low-level,middle-level and high-level features.For the problem of insufficient decision information,a model based on second-order feature fusion method is designed and implemented.The model uses the encoder-decoder structure to provide low-level information,and combines the feature map generated by the encoder with the feature map generated by the deconvolution operation to enhance the feature graph representation ability.The second fusion of feature maps generated by the model in the reconstruction stage to obtain more decision-making information.The experimental results show that the model effectively combines a large amount of effective information in the decision-making stage and improves the recognition accuracy of large objects.