Research On Bridge Detection For Remote Sensing Image

In the 21 st century,with the rapid development of artificial intelligence(AI)technology,rapid and automatic recognition of objects in remote sensing image has gradually become a research hotspot and focus.Along with the rapid growth of aerial remote sensing technology,the explosive growth of the number of optical remote sensing image has provided the necessary dataset support for the use of deep learning method in remote sensing image object detection task.As a typical object in optical remote sensing image,bridge plays an important role in military field and people's livelihood.Based on the research background of automatic detection and recognition of bridge in large optical remote sensing image,this paper studies the topics of feature extraction,road detection,recognition and correction of rotating objects and small object detection etc.According to the characteristics of bridges in remote sensing image – numerous,multi-scale,the bridges are small and their direction is changeable.With the help of powerful feature extraction ability of convolutional neural network(CNN),an automatic recognition and location system of remote sensing image bridge based on deep learning is constructed.The main contents of this paper are as follows:(1)Because of its special shooting angle and large size,the size of the objects in Optical remote sensing image in complex scene is too small,and there are variable scale characteristics between different objects or the same object.In this paper,a new optical remote sensing image cutting method,multi-scale feature extraction and new anchors are introduced,and an improved faster R-CNN object detection algorithm is proposed.The experimental results show that the optimized model can effectively detect the objects in the optical remote sensing image.(2)Based on the above improved faster R-CNN remote sensing image object detection algorithm,this paper studies the multi-scale effect on remote sensing image object detection,and a new method is proposed in order to divide large,medium and small objects in optical remote sensing image,through the analysis of the conventional image scale division standard.(3)On the basis of the above object detection algorithm of remote sensing image,the bridge detection is realized,and a new aspect ratio which is more suitable for bridge is introduced,and the detection accuracy is improved by 1.5%.In addition,through the observation and analysis of datasets and logical inference based on prior knowledge,in general,the appearance of roads is the necessary condition for the appearance of bridges.Therefore,Sobel and Canny operators are firstly used to detect road edge of the dataset to highlight the complete road features,then the detection results fuse with the original dataset to suppress other noise interference,finally,the bridge detection model and the dataset after fusion are used to train and detect in this paper.The experimental results show that the two edge detection operators can achieve the highest accuracy improvement of 1.05% and 0.22% respectively.