Research On Remote Sensing Image Object Detection Algorithm Based On Improved Mask R-CNN

With the development of satellite and remote sensing technology,the amount of remote sensing image data is more and more huge,which is beyond the scope of any manual operation and processing.Therefore,the automatic analysis of remote sensing images plays an important role in urban resource management,search,ocean monitoring,ocean resource management and other practical applications.At the same time,due to the progress of sensor technology,the spatial resolution of the remote sensing image is constantly higher,which makes it possible to locate all kinds of targets(such as aircraft,ships,vehicles,ports,etc.).However,the optical remote sensing image will inevitably be affected by various factors,such as target dense arrangement,similar appearance and small targets,which leads to the degradation of detection algorithm performance,and brings great challenges to the object detection of optical remote sensing image.In view of the problem that the ground targets are densely arranged and the detection boxes are highly overlapped,which leads to the high rate of missed detection,this paper proposes a segment-based object detection algorithm.Firstly,the mask of each target is generated by the bounding box,and the segmentation model Mask R-CNN is trained.In the detection stage,the minimum circumscribed rectangle box of each instance is calculated,so the close bounding box of each target can be obtained.Then,by non-maximum suppression operation on the detection box,the redundant boxes with high overlap can be removed to realize the accurate positioning of the target,effectively eliminating the problem of target missing due to the dense arrangement of the target and the high overlap of the detection box.Experimental results show that the proposed algorithm framework achieves the highest precision.Due to the complex background and similar appearance of some targets,it is easy to cause false detection.Aiming at the problem of complex background and high falsedetection rate of multi class object detection,this paper proposes a class constraint method based on graph convolution neural network.Using the prior information of the objects in the dataset,the relation graph is constructed,and the unstructured information in the relation graph is extracted by the convolution neural network,and the information is fused to the detector,which makes full use of the relationship information between the object categories and corrects the category of false detection,so as to effectively reduce the rate of false detection.Some objects in remote sensing image are small,which are easy to be lost after being sampled by convolution neural network.Aiming at the serious problem of missing small targets,this paper proposes an improved feature fusion method.In the classic FPN structure,the deep features will be continuously fused with the features of the front layers through up-sampling,but the deep feature map has lost most of the information compared with the shallow feature map,so this feature fusion method is a little redundant.Therefore,this paper proposes a more efficient feature fusion method,each layer only uses the features of the adjacent layers,so that the network can focus more on encoding and decoding.At the same time,this paper uses deconvolution as the method of up-sampling on the feature map,which makes the information of the feature map richer and more interpretable.Experimental results show that the improved feature fusion strategy effectively improves the detection rate of small targets,and compared with the classic FPN structure has a significant improvement.In order to verify the effectiveness of the algorithm proposed in this paper,the satellite remote sensing data collected from Google image is used.The experimental results show that the algorithm proposed in this paper has significant advantages in remote sensing image object detection compared with the current advanced algorithms.