Research On The Key Technologies Of Remote Sensing Image Target Detection Based On Deep Learning

Remote sensing image contains abundant ground object information,which can be used in urban planning,agricultural monitoring,ecological services,geological exploration and other aspects.Therefore,the main features of remote sensing images are complex background information,large image size,small target features and dense distribution.Due to the rapid development of remote sensing technology,the technical processing of remote sensing image has higher requirements.Therefore,the remote sensing image target detection technology for rapid and effective extraction of semantic information in remote sensing image has the characteristics of wide scope,high processing efficiency and long operating distance,and has become a key link in remote sensing image processing,and is widely used in all walks of life.With the development of deep learning technology,deep learning has been widely applied in image processing.Its features of automatic identification of image features and accurate positioning and recognition can effectively reduce labor cost and calculation amount,and solve the pain point problem of artificial marking features in the traditional target detection process.Therefore,the combination of deep learning and remote sensing technology has become one of the hot development trends of remote sensing target detection.In this paper,the semantic segmentation and scene classification of remote sensing image target detection are studied.The semantic segmentation model of deep learning is used to extract the regions of interest in remote sensing images,and the scene classification model of deep learning is used to screen the targets in the images,so as to achieve intelligent semantic information extraction of remote sensing images.Specific research contents and conclusions are as follows:(1)For remote sensing images with high resolution,complex information and need semantic segmentation and recognition,this paper builds a network model with coding-decoding structure,proposes a global-local decoder based on Transformer,constructs a feature refinement module to refine the target features of remote sensing images,and uses Dense Net to optimize the decoder.Experiments and ablation experiments on Vaihingen,Potsdam and UAVid data sets demonstrate the effectiveness and efficiency of the proposed method in the application of remote sensing image semantic segmentation.The m Io U and F1 of the model in each data set reached 82.7% and 91.5%,82.5% and 91.3%,77.8% and 92.6% respectively,and F1 was more than 10% higher than other models.At the same time,the segmentation accuracy of small objects(such as cars)is effectively improved,and the semantic information segmentation of ground object boundary is clearer and more complete.(2)For remote sensing images with low resolution,few target feature points in the image,only scene recognition and classification are needed.Since this kind of remote sensing images do not need to extract information and complex model structure pixel by pixel,this paper proposes a lightweight model of remote sensing image scene classification based on improved Dense Net.By combining Dense Net with U-Net,the U-shaped symmetrical structure in U-Net model is used to extract the surface and deep information of remote sensing images in detail.At the same time,ECA attention mechanism and Dense Net jump connection are added,which can not only extract important features related to target features,but also effectively reduce the number of parameters in the model.Drop Block is also introduced to avoid overfitting and improve the robustness of the model.In addition,in order to effectively improve the efficiency and accuracy of classification,a semi-supervised semantic annotation segmentation method is proposed,which can simply mark data and save the cost of manual marking.Through many experiments on UCM data set and Euro SAT data set,the accuracy of the model reaches 96.67% and 99.83%.For the forest and pasture which are easy to be confused,the accuracy rate is more than 90%,and the memory size of the model is only 19.92 MB,which ensures that the accuracy of the model is guaranteed and the lightweight of the model is achieved.Based on the above work,this paper conducts research on remote sensing image target detection,uses the combination of deep learning and remote sensing technology to construct the semantic segmentation model of remote sensing image and the scene classification model of remote sensing image,realizes the location,recognition and classification of target ground objects in remote sensing images,and has good detection effect on multiple remote sensing data sets.It has practical research value and wide application prospect in helping urban development,resource exploration,natural environment protection and so on.