Constrained Convolutional Neural Network For Ship Target Detection In Surveillance Video

Ship target detection is a process that uses algorithms to automatically identify the position and type of the ship in the image.It is widely applied in both civil and military fields.Most of the previous researches are based on high-resolution remote sensing images(radar or near-infrared).With the development of imaging technology,the detection of marine ships based on visible light imaging has attracted the attention of scholars and engineers,and has become a research hotspot in the field of video surveillance.In addition,the rise of convolutional neural networks has greatly accelerated the development of digital image processing technology.Detection algorithms based on deep learning have been widely used in various problems of computer vision.This paper first builds a large-scale,high-quality ship target sample dataset--SeaShips,for training and testing the detection model.Images of the sample dataset come from the surveillance video placed in the Hengqin roundabout in Zhuhai.By extracting frame data,removing redundant images,designing sample diversity,and manually marking process,we get a total of 31455 images of 6 ship types(ore carrier,bulk cargo carrier,general cargo ship,container ship,fishing boat,passenger ship),taking into account the background,illumination,perspective,visible hull part ratio,scale,and occlusion.The comparison between other sample datasets which are published in the academic field(PASCAL VOC2007,CIFAR-10 and Caltech-256)shows that SeaShips is superior to the others in terms of number of pictures,number of targets and sample diversity.Secondly,based on the built SeaShips dataset,this paper completed three basic experiments:(1)Implement three classic detection learning algorithms based on deep learning(Faster R-CNN,YOLO,SSD),and comprehensively summarize the difficulties and challenges of the dataset for ship detection.Comparing the advantages and disadvantages of different detection algorithms to provide basic research results for researchers using the SeaShips dataset;(2)Using the trained YOLO model for actual real-time video surveillance of Hengqin roundabout area in combination with practical applications.Evaluate the effect of real-time detection and realize linkage with UAV images and GIS maps;(3)Implement a ship detection algorithm based on satellite video(top view)and compare it with the video detection effect in(2).Thirdly,this paper studies the new problem of combining deep learning algorithm with traditional algorithm: The deep learning features are abstract and difficult to understand,and the speed of expert knowledge is slow.To solve this problem,this paper proposes a model SFFN that combines expert knowledge features and convolutional neural network features.In the SeaShips dataset,the SFFN model performs well and can effectively improve the detection results under the occlusion of the ship.The first two contributions of this paper have published two papers.The first SCI was published in IEEE Transactions on Multimedia,titled “SeaShips: A Large-Scale Precisely Annotated Dataset for Ship Detection”,I am the second author and correspondent author;The second Chinese core was published in Geospatial Information,entitled "The ship detection algorithm based on YOLOv2 and its application in the monitoring system of Hengqin Island".I am the first author.The third research result is being continuously improved.