Research On Ship Detection Technology Of Wide Visible Remote Sensing Image Based On Deep Learning

With the rapid development of China’s aerospace industry and satellite remote sensing technology,space-based optical remote sensing satellites are developing in the direction of super-resolution,large width and high line frequency.Among them,visible light remote sensing images have become the focus of remote sensing research because of their high resolution,clearer details and texture features,and more intuitive image understanding.However,while obtaining more remote sensing image data resources,it is also accompanied by the decline in the utilization rate of image data,and the traditional image processing technology has been unable to realize the processing of huge data.Marine remote sensing image is an important part of aerospace optical remote sensing image,which contains rich sea information,and ships are an important carrier,so monitoring ship targets plays an important role in military and civil affairs.In this paper,the typical target ship detection in wide visible light remote sensing images is the main research content,and the advantages and disadvantages of different methods in the field of ship detection are investigated.Taking the method based on deep learning as the starting point,in-depth research is carried out around the design of lightweight classification algorithm for remote sensing image scenes,multi-scale ship horizontal frame detection algorithm in complex scenes,dense arrangement and high aspect ratio ship rotation detection algorithm in coastal scenes.The main research contents are summarized as follows:1.For the case that the wide-width image cannot be directly sent to the input end based on the deep learning algorithm,the wide-width remote sensing image is cut into small-size images suitable for the input of the deep learning algorithm and then the subsequent operation is carried out.In view of the fact that there are a large number of non-ship pictures in the segmented small-size pictures,resulting in the waste of detection resources,and the number of parameters of the classification network is far less than that of the detection network,a lightweight remote sensing scene classification algorithm is designed,which classifies the pictures suspected of having a ship from the pictures determined to have no ship,and retains the pictures suspected of having a ship for the next target detection;At the same time,in view of the lack of ship scene classification data set,the RSIS remote sensing scene classification data set covering different shooting scenes,land and sea conditions,weather conditions and interference factors is constructed.The improved Mobile Netv3 algorithm is adopted in the classification algorithm,and the basic convolution is deeply separable convolution.Inverse residual block and linear bottleneck block are used to improve the expression ability of high-dimensional and low-dimensional information of features,SA(Shuffle Attention)module is used to replace SE(Squeeze-and-Excitation)module in the original network,and hard-swish activation function with better optimization effect in the deep network is introduced,which reduces the network parameters and improves the classification accuracy.The FLOPs of the improved Mobile Netv3＿large model is only 233.031 M,and the parameters are only 3.971 M The accuracy of Top1 is 97.5%,the average recall rate is 97.6%,and the average accuracy rate is 97.52%,which achieves a good balance between classification accuracy and model complexity.2.Taking large and medium-sized ships near the coast and small target ships at sea as the research objects,the factors restricting ship detection under multi-scale strong interference are analyzed,and the horizontal frame detection algorithm of IM-YOLOs is proposed.In this algorithm,CA(Coordinate Attention)attention mechanism is integrated into the backbone feature extraction network based on YOLOX-s algorithm without anchor frame,and the fusion mode of feature maps at different levels in the feature pyramid network is improved to be adaptive feature fusion.In the loss function,CIo U loss is used to replace Io U loss,zoom loss is used to replace confidence loss,and the weight of category loss is adjusted to obtain better performance than the baseline algorithm.Among them,the AP50 value reaches 96.77%,the Recall value reaches97.18%,and the FPS is 60.The algorithm can cope with multi-scale changes well,is friendly to small target detection,and achieves a good balance between detection accuracy and detection speed.3.In the dense arrangement scene,the horizontal frame ship detection is easy to miss due to non-maximum suppression,and the background information obtained by the horizontal frame ship detection with high aspect ratio is jumbled,therefore,based on YOLOv5 model,we proposed R-YOLOv5 rotation detection algorithm.With CSPDark Net as the feature extraction network,PANet as the feature pyramid network,and Si LU activation function,angle prediction branch is added on the basis of YOLOv5’s original algorithm,and the angle regression problem is transformed into a classification problem.The predefined angle is discretized by CSL(Circular Smooth Label)method,and the loss about angle classification is added to the total loss function.In the HRSC2016 data set,the final AP50 value of R-YOLOv5 m reached 90.31%,and the FPS reached 49.The designed R-YOLOv5 algorithm achieved a good balance between accuracy and speed.Through the detection and analysis of test set pictures,RYOLOv5 can effectively solve the problems of missing detection of densely arranged ships and poor fitting degree of high aspect ratio ship detection frames.To sum up,in this paper,a variety of scenarios and technologies involved in the ship detection task of wide remote sensing images based on deep learning are deeply explored,and some new ideas and improvement methods are put forward,focusing on the balance between model accuracy and speed index.The method and conclusion proposed in this paper can provide reference for the ship automatic detection system with wide remote sensing images.