The Method Of Typical Object Detection In Border And Coastal Defense Based On Deep Learning

Our land area is huge,with extremely long border and coastline.Border defense takes on the important mission of protecting the homeland and protecting our country.At present,a large number of video capture devices such as cameras are deployed in border areas of the PRC,and the collected video are manually processed by these border guard,which is extremely inefficient and can't meet the requirement of high detecting accuracy of border defense.Using the traditional object detection method of computer vision to process video data depends too much on prior knowledge.Meanwhile,the robustness of the method is not high and cannot satisfy the required detection accuracy of border defense when the target deforms,ambient lighting changes and so on.At present,the development of big data,cloud computing and artificial intelligence technology,especially the improvement of automatic object detection and recognition technology based on deep learning algorithm,make it possible to use deep learning methods to detect the objects in the border and coastal defense images.In this paper,the study of object detection in border and coastal defense images based on deep learning is carried out.The main work is showed:(1)Firstly,the object detection based on deep learning and its related technical theories are reviewed,and the object detection framework which is called Faster RCNN based on candidate regions is introduced systematically.(2)An improved object detection method based on candidate regions is proposed.First,the SE module is added to the original classification network to improve the correlation between convolutional channels.And deformable convolution is used to replace the original special convolution layer in feature extraction structure,to improve the ability for adapting to deformation and different scales of the object,eventually to achieve the goal of higher accuracy.Secondly,using a combination of Inception structure,the depthwise separable convolution and the position-sensitive score map which is used to replace the fully connected layer of original detection network,to achieve real-time object detection performance.And the experiments show that the detection algorithm in this paper,compared with the existing Faster RCNN detection model,can greatly improve the object detection speed and correspondingly increase the accuracy.(3)The improved object detection method is verified by experiments and simulation.The mean average precision is compared on the PASCAL VOC2007 and VOC0712 datasets.And the detection speed is compared on pictures of different resolution.Finally,the generalization ability of network is analyzed and verified by experiments on self-built data sets.The experimental results show that the proposed method is superior to the current methods which meet the requirements of real-time and accuracy in the applications of frontier and coast defense.And the method has a good generalization ability in practical scene applications.(4)After completing the above theoretical method improvement and experimental simulation,a visual prototype platform based on deep learning is designed and implemented through different application scenarios by using the combination of QT,Python,Shell,Mxnet,C++ and other development tools and languages.