Research On The Improvement Method Of Remote Sensing Image Ship Target Detection Based On Deep Learning

The resolution of remote sensing images has increased with the maturation of satellite technology.The ship detection technology based on remote sensing images makes it possible to monitor a large range and far sea area,which can greatly enrich the monitoring means of maritime departments.Satellite remote sensing technology is maturing,resulting in higher resolution of images and the proliferation of data.There are disadvantages to traditional ship detection algorithms,including low recognition accuracy,low efficiency,and susceptibility to interference from the background,which are difficult to meet the application requirements.Convolutional neural network has been extensively applied in the domain of object detection because of its powerful feature extraction ability.Object detection algorithm based on deep learning is a cutting-edge technology to solve the problem of ship target detection and classification recognition.This paper focuses on the task of ship target detection for high-resolution remote sensing images,and the main work is as follows.(1)A high-resolution ship remote sensing image dataset was constructed to solve the problem of the lack of small ship samples in the existing dataset.Targets were labeled into four categories,including aircraft carriers,battleships,submarines and merchant ships,with resolutions ranging from 0.4m to 2m,and the number of images was 1333.A total of 4969 ship targets were annotated,including 163 aircraft carriers,2285 battleships,2106 merchant ships and 415 submarines.(2)An anti-interference model YOLOv4-CR was proposed for small target detection.The m AP value of YOLOv4 model was only 77.66% when trained on the self-built dataset,and it missed small targets and was difficult to resist complex background interference.The network structure was improved to address the problems of poor small target detection and poor interference resistance of the YOLOv4 model:the RFB＿s module was introduced to improve the acceptance domain to improve small target detection;the CBAM double attention mechanism was introduced to enhance the difference between target and background,improve target saliency,improve small target detection,and enhance the ability to resist complex background interference.The Kmeans++ algorithm was also introduced to re-cluster the targets to obtain an a priori anchor frame suitable for the ship,and the YOLOv4-CR model was trained by loading the parameters of the backbone part of the pre-trained model instead of random initialization.The m AP value of the YOLOv4-CR model improved to 91.40%,and the small target detection effect and anti-interference ability performed better.However,the detection speed was reduced.(3)An anchor-free frame detection model YOLOv4-F was proposed to solve the problem of slightly lower processing speed of YOLOv4-CR.The backbone network was improved,and decoupled detection head was used to predict regression and classification respectively,which learns features that are beneficial to regression and classification separately.The anchor-free detection method and matching positive samples with the sim OTA strategy were used to reduce the number of parameters output by the network.Then the anchor-free model was trained,and the results showed that the m AP value of the anchorless detection model YOLOv4-F improved to 93.16% and the FPS value improved from 20.34 to 24.85,which was 2.52 higher than the FPS value of the YOLOv4 model;and the detection effect of the YOLOv4-F model on the dense adjacent row of ship targets improved,and the small target detection and antiinterference capability was also further improved.