Object Detection Of Remote Sensing Image Based On Deep Convolutional Neural Network

Object detection of remote sensing images is widely used in disaster monitoring,military reconnaissance,and urban planning.It is a hot research issue in the field of aviation and satellite image analysis,and it is also the basis of many remote sensing image analysis tasks.With the development of remote sensor technology and the increase in the number of satellites,the spatial resolution of optical remote sensing images has gradually improved,providing more and more detailed information,making it possible to identify the object or specific object categories of remote sensing images,the results are more intuitive to a human decision maker.Therefore,object detection and recognition based on optical remote sensing images are becoming more and more important.In recent years,the application of deep learning methods to object detection in natural scenes has developed rapidly,and researchers have begun to try to combine deep learning methods with remote sensing image object detection.Although the remote sensing image and the natural scene image contain similar image features,but unlike the close-range imaging of the natural scene image,the remote sensing image has the characteristics of uncertain object direction,large change in object scale,dense object distribution,and complex background.The object detection method in the natural scene is applied to the object detection of remote sensing images,and it cannot achieve satisfactory results.Therefore,this paper focuses on the application of deep learning in remote sensing image object detection.The main research contents include::(1)Aiming at the difficult problem of complex background and large scale change of remote sensing image target,a multi-scale remote sensing image target detection method based on attention mechanism is designed.First,through the improved balanced feature pyramid network,the feature maps used in each layer of prediction are fused with different resolutions and different semantic information.The fused feature maps are used for object detection of different sizes,and the model is improved for object of different scales.Then,through the saliency detection module based on the spatial and channel attention mechanism,the feature extraction network is different for the key feature areas that different objects pay attention to,to obtain the target area of focus,and improve the model to the object in a complex background.Finally,the effectiveness of the algorithm is verified on the public data sets DOTA and NWPU VHR-10.(2)Aiming at the difficult problem that the objects in the remote sensing image have dense distribution and uncertain direction,a remote sensing image object detection method based on rotating area is designed,which can be used for rotating object detection,especially ship object detection.The network uses rotating proposal to represent object detection results,and further identifies the ship's model.Through experimental analysis,this representation method is superior to the existing horizontal proposal representation method.In order to achieve the object detection based on the rotation proposal,the region proposal network has been improved,including rotating the region of interest pooling layer and rotation region regression,so that it can output the proposal region with rotation angle.Finally,by introducing positioning accuracy to guide the non-maximum suppression algorithm,the post-processing algorithm is optimized.The effectiveness of the proposed algorithm is proved by experiments on the public data set HRSC2016..