Research On Remote Sensing Ship Recognition In Complex Scene Based On Capsule Network

Remote sensing image detection is an important research topic in computer vision tasks.It is the most typical example of a significant scene target detection.It plays a vital role in sea area security monitoring,port ship detection,and moving target tracking.With the development of target detection technology,more and more new detection methods have shown their superiority in various fields.The application of target detection technology in remote sensing ship detection tasks not only promotes the solution of problems in remote sensing image recognition but also enables breakthroughs in target detection technology in practical applications.This article mainly accomplishes the following work in this regard.(1)Aiming at the recognition difficulties of cloud occlusion,shore-based interference,and small targets in remote sensing images,based on the more effective Mask R-CNN in the current object detection algorithm,and make corresponding improvements.A residual visual attention mechanism was added to the feature extraction network to improve the network's ability to extract ship feature information and solve the problem that the network could not extract enough ship feature information in remote sensing images with cloud occlusion.Multi-level features extraction and dequantization operation methods in Mask R-CNN could solve the problem of missed detection of small targets.A re-learning strategy of difficult samples is used to enhance the model's ability to detect complex scenes.Through the above methods and strategy,the overall accuracy rate of ship detection reached 92.56%,and the recall rate reached 89.26%.(2)The capsule network is studied.While conducting fundamental research on the capsule network,the model functions are expanded to enable the capsule network to perform single target detection and apply it to remote sensing ship target detection.In the case of insufficient capsule vector representation capabilities,the ship target classification and ship target positioning functions were separated by a dual-branch capsule structure,and the ship classification and ship positioning were achieved without using traditional violence search to obtain the target area.(3)Combining the advantages of Mask R-CNN and capsule network in ship detection,a new ship detection network-Caps-RCNN is constructed.This method retains the RPN and the target regression part in the detection network in Mask R-CNN.Use these two parts to complete the search of the target position in the remote sensing image.After that,use the capsule network to classify the ships in the searched target area and give the probability of ship existence.Then,use the probability of ship existence to suppress the target area output by the RPN and shield the non-ship proposal box.Without considering the efficiency of the algorithm,this method can improve the detection result and better improve the missed and misdetected situations of Mask R-CNN in ship detection.Through the above improvements,the overall accuracy rate of ship detection reached 97.37%,and the recall rate reached 96.26%.