Semantic Description Of Remote Sensing Images Based On Deep Learning

With the improvement of modern remote sensing technology,how to extract valuable information from massive remote sensing images has become an important research topic in the field of remote sensing.At present,the interpretation of high resolution remote sensing scenes is mostly focused on the object oriented classification,but there is still a"semantic gap" between the classification results and the semantic information of the high-level scene.Therefore,the intelligent translation of high resolution remote sensing images from different levels and different angles has become the most challenging scientific frontier in the field of remote sensing,and has important scientific significance.In addition,remote sensing image content understanding and description can provide decision level support for remote sensing applications,and has wide practical application value.This paper fully excavate the structure characteristics of high resolution remote sensing image space,scene space and semantic space,and combine the Natural Language Processing technology to realize the understanding and description of the high resolution remote sensing image content.The contents of this paper are as follows:(1)a remote sensing semantic description method based on fast regional convolution neural network is proposed,and a fast regional convolution neural network is constructed to extract the visual features of high resolution remote sensing images,and a bidirectional recurrent neural network is constructed to extract text features,and the image features and text features are unified into the same dimension space.In contrast,the high and low resolution remote sensing image can be described by the long and short time memory model.The experiment shows that the method fully takes into account the rich information information of the high resolution remote sensing image and the complex and diverse features of the scene.It combines the target detection method to extract the effective visual features,and then generates a precise and diverse semantic description of the text.(2)the remote sensing semantic description of the convolution recurrent neural network based on local response is presented,and the "visual attention mechanism" is introduced.When the text is generated,not only the problem of target detection is considered,but also the relationship between the targets is taken into account,and the long and short-term visual information and text information are stored in implicit memory.Optimize network structure and finish end to end training.The experiment shows that the method takes more full account of the position correlation of the target in the visual information and memory text context,and then generates more accurate and rich text semantic description.(3)a semantic description method of remote sensing based on semantic embeddedness generation network is proposed,a regional convolution neural network for remote sensing image target detection task is constructed,and a remote sensing text analysis model based on long and short time memory network is constructed.Network modeling is used to analyze the correlation between visual information and text information,and then generate text sentences to achieve the understanding and description of high resolution remote sensing scene content.The experiment shows that the method fully considers the features of high resolution remote sensing images with large scale changes and small number of samples,and excavate the visual information in the image and generate accurate and rich semantic description of the text.