Optical Remote Sensing Image Object Detection Based On Multi-scale Feature Fusion And The Prediction Of Oriented Bounding Box

Object detection is an important research area in remote sensing field,which shows significant potential applications for military and civilian areas.With the rapid development of deep learning,Convolutional Neural Networks(CNNs)have made great progress in natural scene image classification and object detection tasks with their powerful feature extraction capabilities,and have been widely used in remote sensing image object detection tasks.Due to the drawback of low accuracy caused by the wide range of object scales,small size object,complex background and densely distributed objects in optical remote sensing image object detection task,this thesis proposes an optical remote sensing image object detection method based on multi-scale feature fusion and the prediction of oriented bounding box.The main contributions of this thesis are listed as follows:1? Due to the low detection accuracy caused by the wide distribution of the object scale in remote sensing image,this thesis proposes an object detection method in optical remote sensing image based on the improved SSD.Firstly,multiple detection layers are set at different levels of the network to detect objects with different scales.Secondly,for the designing of the detection structure,the fusion result of the feature maps obtained from multiple sets of convolutions are used to predict the position and category of objects,which improves the adaptability of the network to different scale objects.The network is more suitable for the detection of remote sensing image objects by using the K-Means clustering algorithm to analyze the object scale characteristics of the data set as a reference for the default box design.And the design of a multi-granularity detection strategy in the test process has further improved the detection accuracy by combining the detection results with different degrees of refinement.2? Due to the low detection accuracy caused by the small object size and the high false alarms resulted from the complex background,this thesis proposes a Coarse-to-Fine Ship Detection Network(CF-SDN)for ship detection in optical remote sensing images.CF-SDN sets multiple detection layers at different levels of the network and adopts a feature extraction structure that combines different hierarchical features to improve the representation ability of features,which enhances the semantic information of the small objects and improves the detection accuracy of small objects.A coarse-to-fine detection strategy is used at each detection layer,which adds a fine classification network based on the traditional classification network and the bounding box regression network,and the false alarms generated in the previous detection are removed by further fine classification.In order to avoid the false alarms on land in the test process,this thesis designs a robust sea-land separation algorithm which comprehensive accounts of the gradient information and gray information.The sea-land separation algorithm adapts to the optical remote sensing image with complex situations and improves the accuracy and efficiency of ship detection.3? Due to the detection difficultly of densely distributed objects,this thesis proposes a ship detection method based on the prediction of the oriented bounding box(OBB)in optical remote sensing images.On the basis of CF-SDN in Algorithm 2,a set of anchors with multiple directions are set as object candidate boxes at each detection layer,while the regression parameter of the detection layer's bounding box regression network adds the direction information of the object.We introduce the method of calculating the intersection-over-union of the OBB in the network.The network finally predicts some oriented bounding boxes.Compared with the traditional bounding box,oriented bounding box describes the position of the object more accurately by increasing the direction information of the object,which benefits the detection of the densely distributed object.