Research On Sea Surface Object Detection And Movement Prediction Of Satellite Image Based On Deep Learning

The sea object detection task based on satellite images requires marking the position of the object in the image,and accurately giving the category of the object.The current object detection methods are divided into two main directions:two-step detection method and one-step detection method.The advantage of two-step detection method is high accuracy,but the disadvantage is slow speed.One-step detection method has the advantage of fast detection speed and the disadvantage of poor accuracy.This paper aims at the application research of ship object detection in satellite image,and the task real-time requirement is high.Therefore,this paper studies the one-step detection algorithm.One-step detection algorithm is an engineering application friendly method,but its low accuracy is still the key factor restricting its large area application.In order to improve the accuracy of this type of target detection method,the high-level bidirectional multi-scale feature fusion module and generalized intersection ratio loss function are adopted in this paper to greatly improve the recall rate and slightly improve the average accuracy of the category.In addition,the deep separable convolution is used to replace the ordinary convolution,which alleviates the computational speed decline caused by the addition of the bi-directional multi-scale feature fusion module at the high level.In order to maintain a faster computation speed,the number of channels in the feature extraction network is compressed in combination with the small number of ship classes in this paper.In order to maintain a faster computation speed,the number of channels in the feature extraction network is compressed in combination with the small number of ship classes in this paper.On the surface of the experimental results,the improved high-level bidirectional multi-scale feature fusion module is helpful to improve the detection performance.This paper also explores the feasibility of ship movement prediction with the help of Moving-MNIST data set.Ship motion prediction belongs to the problem of image sequence prediction.The prediction result map has the disadvantages of pixel blur and spatial structure information loss.This paper proposes a multi-branch convolutional long-term and short-term memory neural network,which effectively enhances the spatial structure information expression ability of the sequence image prediction results,and the pixel definition has been greatly improved.